VKD3D-Proton 3.0 brings FSR4 support, shader backend rewrite, lots of bug fixes and performance upgrades

A new major release, yay!
A few milestones have been reached over the last year, warranting a new major bump.
It's been quite a while since the last release due to new things coming up constantly.
These tags are mostly arbitrary anyway, and tend to be done when islands of calm and stability emerge.

Major items

DXBC shader backend rewrite

@doitsujin rewrote the entire DXBC backend, replacing our legacy vkd3d-shader path.
DXVK and vkd3d-proton now share the same DXBC frontend which gives us clean,
"readable" (as readable as DXBC can be) and lean IR to work with.
dxil-spirv standalone project now supports DXBC as well as a result.

Lots of games which used to be completely broken before due to bugs and missing features
in the legacy vkd3d-shader backend are now fixed. E.g. Red Dead Redemption 2 runs just fine now in D3D12 mode.
Some recently released DXBC based games also only work on the new path.
The amount of regressions found the last months in DXBC games has been very minor,
but it's possible there are still bugs in this area.
However, given that DXVK uses it now as well, it's been battle tested quite extensively already.

FSR4 support

We added support for AGS WMMA intrinsics through VK_KHR_cooperative_matrix and VK_KHR_shader_float8,
which is enough to support FSR4.
Note that these shaders are tightly coded for AMD GPUs with some implementation defined behavior
(particularly around matrix layouts), and they will not necessarily work on other GPU vendors.

There is also a quite hacky emulation path of this which relies on int8 and float16 cooperative matrix support,
which can run on older GPUs at significant performance cost (and some cost to theoretical correctness).

Note that the default "official" build of vkd3d-proton only exposes this feature when the native
VK_KHR_shader_float8 is properly supported, i.e. RDNA4+ only.
The emulation path is available when building from source with the appropriate build flags.
The decision to not include this emulation path by default is over my pay grade.
The aim is to be able to ship FSR4 in a more proper way in Proton.

Features

We've more or less caught up on the things we can feasibly implement,
so there isn't much exciting stuff happening on the feature front.

• Implemented experimental support for D3D12 work graphs. No real-world content ships this yet.
  This implementation is far from complete,
  but it works on "any" GPU since we emulate the feature with normal compute shaders.
  Funnily enough, the performance of this emulation can massively outperform native driver implementations of the feature
  in many scenarios we've tested (at the cost of some extra VRAM usage).
  See docs/ for more details on implementation and some performance numbers.
• Expose AdvancedTextureOpsSupported by default from SM 6.7 if VK_KHR_maintenance8 is supported.
• Expose the recently added sparse TIER_4.
• Bump exposed D3D12SDKVersion to latest 618.
• Experimentally expose support for opacity micromaps.
  There are some details which aren't quite compatible with the D3D12 API, but some basic demo content is working fine.
• Add support for AMD_anti_lag when exposed. The current implementation does not take frame-gen into account.
• Implement support for tight alignment from recent AgilitySDK.
• Add support for shared resource path on upstream Wine.

Performance

• Overhaul the texture copy batching situation.
  The new batching logic should be able to improve performance in many more cases than before.
  • Implemented support for VK_KHR_unified_image_layouts.
    Image copy batching in particular can take advantage of this to avoid a lot of unnecessary barriers.
• Removed manual clear workaround on newer (6.15.9+) kernels on AMD, where an old kernel regression was finally fixed.
  Kernels older than 6.10 are also not affected by this workaround.
• Use push descriptor path on Qualcomm GPUs over BDA for speed.
• Improve handling of GDeflate when decompression extension is not available.
  We now ship our own fallback shader in GLSL instead of the more awkward HLSL shader that dstorage ships.
• Bump DGC scratch size on NVIDIA. Should avoid some massive perf drops in Halo Infinite on NVIDIA.
• Add performance optimization for The Last of Us Part 1 to prefer 2D tiling on 3D images.
  Requires an update to Mesa as well to get the proper effect.
• Handle depth/stencil <-> color image copies better when VK_KHR_maintenance8 is supported.
• Make use of VK_EXT_zero_initialize_device_memory to avoid manual clears on allocation.

Fixes

• Emit render pass barriers as expected on tiled GPUs. Fixes misc rendering bugs reported on e.g. Turnip.
  • For performance reasons, we deliberately skirt the spec a bit on desktop GPUs.
• Fixed a bunch of minor correctness problems exposed by new Vulkan-ValidationLayers.
• Adjust how PointSamplingAddressesNeverRoundUp is reported to match recent driver behaviors.
• Fix overflow bugs in massive (> 4GiB) sparse resource handling.
• Fix reporting of some esoteric format properties to better match native drivers.
• Fix handling of NULL acceleration structure descriptors.
• Fix some texturing bugs in Helldivers II on NVIDIA.
• Fix some bugs with memory type handling on very old NVIDIA GPUs.
• Fix bug when pixel shader includes root signature.
• Make ClearUAV barrier insertion the default now.
  Too many games screw this up, and D3D12 drivers seem to do it by default.
• Fix shared fences when initial value is not 0. Fixes some Star Citizen issues.
• Fix rare deadlock scenario in Ninja Gaiden 4.
  Fixes some long-standing issues with how we deal with fence rewinds.
• Fix some long-standing issues with how we deal with placed MSAA resources and alignment.
• Make sure we don't clear memory of imported resources.
  This doesn't fix any known games, but you never know :V
• Improve correctness for many odd GS/HS/DS corner cases with primitive types and API validation.
• Fixes crashes when index buffer SizeInBytes = 0, but VA was invalid.
  Seen in some Saber Interactive games.
• Fixes some potential deadlocks in VR interop APIs when multiple threads attempt to acquire Vulkan queue.
• Fixes 16-bit aligned structured buffer strides. Not observed in any real content, but you never know!

Workarounds

• Add FF VII rebirth sync bugs workarounds. Fixes some rare GPU hangs.
• Add misc AMD workarounds for Monster Hunter Wilds caused by bugged hardware around sparse SMEM.
  • A proper hardware workaround in RADV is still pending.
• Workaround some Starfield bugs around NonUniformResourceIndex use.
• Add performance workarounds for extremely large tessellation factors used in misc new Koei Tecmo games.
• Add Wreckfest 2 workarounds for illegal texture placement aliasing. Fixes some broken textures.
• Add barrier in Satisfactory that game missed. Fixes some corrupt rendering especially on AMD.
• Ignore NOT_CLEARED flags on allocation in all games now. Native drivers seem to always clear regardless of the flag,
  and e.g. Street Fighter 6 relies on NOT_CLEARED memory to actually be cleared :(
• Workaround some issues with RGB9E5 and alpha write masks observed in Ninja Gaiden 4.
• Add missing barrier in Death Stranding (the older build, not Director's Cut).
• Add missing barrier in Wuthering Waves.
• Workaround bugged uninitialized loop variable in Dune MMO.
• Disable UAV compression in Spider-Man Remastered. Fixes some weird RT issues on RDNA2.
• Add Root CBV robustness workaround for Gray Zone Warfare.
• Disables color compression in Rise of the Tomb Raider. Fixes some glitches due to game bug on AMD.
• Workaround some bugs in Port Royal benchmark.
• Workaround Mafia: Definitive Edition hanging GPU when using FSR on startup due to use-after-free.
  • The workaround applies to all uses of FSR. Plausibly workaround a hang in MGS: Delta as well, but not confirmed it was this bug.
• Workaround Control RT path occasionally observing NaNs due to bad normalize() patterns.
• Workaround Final Fantasy Tactics Ivalice Chronicles illegally using dynamically indexed root constants.

Misc

• Added a lot more debug instrumentation as usual.
  • Not user facing, so omitting details.
• Make it a bit easier to use vkd3d-proton in Linux-native projects.
• Remove DXVK_FRAME_RATE to align with DXVK's removal. Only VKD3D_FRAME_RATE remains (at least for now).