LinAlg HLK: Add bfloat16 coverage once DXIL gains a bfloat16 ComponentType

Goal

Extend the LinAlg CPU oracle and the DxilConf_SM610_LinAlg execution tests to cover bfloat16 matrices, once the prerequisites below have landed.

Why

hlsl-specs#907 adds a bfloat16 format to HLSL proposal 0035. The LinAlg CPU oracle added in #8666 models F16, F32, I32 and U32 only, so bfloat16 matrices cannot currently be constructed, encoded or compared against an independent host reference.

#907 defines bfloat16 as "an optional feature that hardware vendors can choose to support", so coverage must be capability-gated rather than unconditional. An implementation that does not support bfloat16 must not fail conformance for it.

Blocked on

• hlsl-specs#907 merges.
• DXIL ComponentType gains a bfloat16 member. include/dxc/DXIL/DxilConstants.h currently ends at F8_E5M2 = 22, so there is no enumerator for the oracle to switch on.
• A host-side bfloat16 type exists in tools/clang/unittests/HLSLExec/HlslTestDataTypes.h. There is no equivalent of HLSLHalf_t for bfloat16 today.

Required work

• Add a bfloat16 alternative to cpu_oracle::TypedMatrixValues and a matching ComponentTraits specialisation. TypedMatrix::compType() derives the component type from the active alternative, so there is no separate tag to update.
• Add bfloat16 arms to isSupportedComponentType, componentTypeName and makeSequentialMatrix, including a representable-value range check equivalent to the existing F16 and F32 checks.
• Add a bfloat16 case to elementSize.
• Gate bfloat16 tests on the reported capability so unsupported implementations report not-applicable rather than failed.
• Add execution test coverage for the operations that accept bfloat16.

Acceptance criteria

• The oracle can construct, encode, decode and compare bfloat16 matrices against hand-derived expected values.
• Expected values come from an independent host reference, not from the shader under test.
• Implementations without bfloat16 support report not-applicable, not failed.
• Coverage matches the operation set that hlsl-specs#907 permits for bfloat16.

Public references

• Raised in review of #8666: https://github.com/microsoft/DirectXShaderCompiler/pull/8666#discussion_r3685978024
• Parent tracker #7841
• hlsl-specs#907
• LinAlgTests.cpp

Out of scope

• Adding the bfloat16 enumerator to DXIL, or any compiler-side bfloat16 support.
• Expanding non-bfloat16 type coverage.

Assisted-by: GitHub Copilot

After hlsl-specs#907 and the prerequisite DXIL and host-type changes land, read tools/clang/unittests/HLSLExec/HlslTestDataTypes.h, the CPU oracle helpers, and LinAlgTests.cpp; check DxilConstants.h for the new component type. Run the LinAlg execution tests and verify construction, encoding, decoding, independent expected values, capability-gated not-applicable results, and permitted operation coverage.