[HLSL] Add non-square LinAlg CopyConvert coverage

[JoeCitizen]

Add non-square SM6.10 Linear Algebra CopyConvert execution coverage.

• Adds CopyConvert_Wave_4x8_F32_Transpose using independently derived expected values.
• Parameterizes source and destination dimensions and row strides for CopyConvert.
• Corrects CPU-reference indexing that was previously masked by square-only matrices.
• Preserves the existing 16x16 F16 CopyConvert behaviour.

Validation

• Built ExecHLSLTests.dll, dxcompiler.dll and dxc.exe successfully.
• The new test passed end-to-end on Microsoft Basic Render Driver using WARP 1.65535.20-preview and D3D12 Agility SDK 1.721.2-preview.
• Both existing 16x16 F16 CopyConvert tests passed on the same runtime.
• Full LinAlg class: 20 total, 19 passed, 0 failed and 1 pre-existing OuterProduct skip because the released SDK headers do not expose DIRECT3D_LINEAR_ALGEBRA.
• No physical GPU or HLK lab execution is claimed.

AI assistance and human review

This draft contains substantial agent-generated content and is intentionally not ready for maintainer review. Before marking it ready, a named human must read every changed line, independently verify the expected 8x4 transpose and source/destination strides, understand the implementation, and own the final rationale.

Assisted-by: GitHub Copilot