[HLSL] Add LinAlg MatVec interpretation and bias coverage

[JoeCitizen]

Summary

• refactor the MatVec execution harness to use independent matrix, vector, bias, and output resources with exact typed host encoding and CPU-derived dot-product expectations
• add exact ThreadVectorMatrixMultiply capability queries and cover non-uniform F16, capability-gated ColumnMajor, required packed SInt8 and UInt8 tuples, the required native F32-to-SInt8 interpretation path, optional unsigned U32 output, and independent bias data
• correct packed 8-bit element sizing and validate LSB-first vector packing with zero padding to the 32-bit storage boundary
• separate native F32 source values from their hand-derived SInt8 interpretations, covering RTNE ties and positive/negative saturation, and use high-bit UInt8 lanes that distinguish unsigned from signed packed interpretation

The four existing MatVec baselines are migrated to the shared harness without changing their intended inputs or results.

Validation

• built the Release ExecHLSLTests target
• ran all 11 MatVec methods on the compatible local WARP and Agility SDK 1.721.2-preview: 10 passed and the ColumnMajor case skipped because WARP does not advertise TRANSPOSE
• after propagating the correction through the complete stack, all 66 LinAlg methods reported 61 passed, 0 failed, and 5 authoritative capability-backed skips
• compiled the capability-gated ColumnMajor shader independently
• inspected emitted DXIL for packed SInt8/UInt8 inputs, native <8 x float> input with SInt8 interpretation, unsigned output with isOutputSigned=false, and a distinct bias SRV
• compiled HlslExecTestUtils.cpp and LinAlgTests.cpp against the preview D3D12 headers
• clang-format 17.0.1 and git diff --check pass

No physical GPU or packaged-HLK execution is claimed.

Preview WARP defect

Public preview WARP advertises the mandatory F32-vector x SInt8-matrix to SInt32 tuple as supported, but the strengthened test receives [-372, 372, -130, 262] instead of the independently derived [3336, -3336, -2, -254]. The physical inputs include half-way values and values beyond the SInt8 range, so the case now directly exercises RTNE ties and positive/negative saturation.

The emitted DXIL correctly carries a native <8 x float> vector and SInt8 conversion-target interpretation. The current WARP implementation treats the float words as packed I8 bytes instead of applying the required RTNE saturating F32-to-SInt8 conversion. The mandatory test remains active; skipping or weakening it would hide a Tier-1 conformance defect.

Stack

This draft is stacked on PR #8671, which is stacked on PR #8670, PR #8669, PR #8668, PR #8667, PR #8666, PR #8665, and PR #8662. Until those ancestors land, this diff contains their commits as well. The implementation commit is f989f89a9; review correction 247f38d10 strengthens the F32-to-SInt8 and UInt8 interpretation oracles.

This remains a draft for named human review. The reviewer should verify the capability-query tuple, native-vector interpretation immediate, RTNE/saturation hand derivation, high-bit UInt8 packing, independent bias resource, CPU oracle, and preview-WARP defect analysis before requesting maintainer review.

Refs #7841
Refs #8559
Refs #8560
Refs #8650
Refs #8653

Assisted-by: GitHub Copilot