RTAO: fix denoiser's unsafe wave-lane ↔ SV_GroupIndex ordering assumption

[amarpMSFT]

What

The RTAO denoiser's CalculateMeanVarianceCS and DisocclusionBlur3x3CS compute shaders exchanged each row's values across wave lanes while deriving source-lane indices from SV_GroupIndex. That assumes row-major thread-to-lane packing and a wave width of at least 16, neither of which HLSL guarantees for these 2D thread groups.

This PR removes that assumption. The affected shaders now exchange values only through groupshared-memory caches indexed by their logical thread coordinates. There is no optional wave path or runtime capability gate.

Why it's a bug

Both live shaders use 2D thread groups ([numthreads(8, 8, 1)]) and previously did the equivalent of:

const uint Row_BaseWaveLaneIndex = (WaveGetLaneIndex() / 16) * 16;
...
float cValue = WaveReadLaneAt(value, Row_BaseWaveLaneIndex + /* column derived from SV_GroupIndex */);

This returns the intended neighbour only if the implementation maps each thread's wave lane index to its row-major SV_GroupIndex. The HLSL SM 6.6 derivatives specification explicitly states that for 2D thread groups "no relation between the values of SV_GroupIndex and the return value of WaveGetLaneIndex() should be assumed." Wave-op support and wave width cannot establish that ordering.

Repro

On WARP (wave width 4), the old shader path reads nonexistent or unrelated lanes. A D3D12 compute harness ran the affected shaders against CPU references:

• CalculateMeanVariance: the old wave path was wrong on 576/576 tested interior pixels. The groupshared implementation matched to half precision (maximum mean delta 0.000338; maximum variance delta 0.000296).
• Depth-aware gaussian/disocclusion filtering: the old wave path was wrong on 35/36 tested interior pixels. The groupshared implementation matched the CPU reference.

The fix

• CalculateMeanVarianceCS.hlsl now loads the logical 16x16 input tile into groupshared memory, synchronizes the group, and calculates each horizontal kernel from that cache.
• DisocclusionBlur3x3CS.hlsl now does the equivalent groupshared exchange for each (value, depth) pair before applying the depth-aware gaussian weights.
• The same latent pattern was corrected in the currently-unused depth-aware separable gaussian variant. The file was renamed to DepthAwareSeparableGaussianFilter3x3CS.hlsl because it no longer uses WaveReadLaneAt.
• Obsolete wave-support requirements and stale measured/relative performance claims were removed from the README and affected shader comments.

Validation

• All three final shaders compile successfully with dxc as cs_6_3.
• Their final DXIL is byte-identical to the groupshared binaries exercised by the WARP CPU-reference A/B tests above.
• A source scan confirms that no WaveReadLaneAt call or deleted fast-path identifier remains in this sample.
• The final PR adds no C++ or project-system changes; those files are byte-identical to origin/master.
• git diff --check passes, and the project XML files parse successfully.

Notes for reviewers

The earlier capability-gated wave permutation has been removed rather than retained behind WaveOps/WaveLaneCountMin: those feature values describe intrinsic availability and width, not lane-to-thread ordering. Groupshared exchange is now the sole implementation, so correctness is independent of wave size and lane assignment.

[Jesse Natalie (jenatali)]

LGTM