[HLSL] Correct LinAlg OuterProduct accumulator use

[JoeCitizen]

Summary

• Declare the OuterProduct result with the shared USE parameter instead of the legacy hard-coded A use.
• Set the host parameters to MatrixUse::Accumulator, matching proposal 0035 and <dx/linalg.h>.
• Assert the accumulator-use invariant in the host runner.

This is a narrow baseline correction; it does not add new shapes, types, scopes, or capability handling.

Validation

• clang-format 17.0.1 and git diff --check pass.
• ExecHLSLTests builds successfully.
• Standalone DXIL inspection changes the matrix type from LinAlgMatrixC8M16N16U0S0 to LinAlgMatrixC8M16N16U2S0; OuterProduct and descriptor accumulation consistently use the accumulator type.
• The focused TAEF test executed end-to-end against preview D3D12 headers, Agility 1.721.2-preview, and WARP 1.65535.20-preview: 1 passed, 0 failed, 0 skipped. The exact value-4 readback oracle passed.
• The currently published preview WARP has the known ConvertLinearAlgebraMatrix compute-command-list bug documented in [SM6.10][Exec][Bugfix] Add matrix layout conversion helpers, add post-dispatch callback, and re-enable OuterProduct test #8606: recording the conversion on a compute list poisons Close() with E_FAIL. For local validation only, the main ShaderOp command list was temporarily changed to DIRECT, which the API specification permits. No queue workaround is included in this PR. Unchanged upstream/main@2fab3f175 and this branch fail identically on the affected compute-list runtime; this branch passes on the documented DIRECT-list workaround.

Refs #7841
Refs #8563
Refs #8561

Assisted-by: GitHub Copilot