Needs to check if triton LLVMToSPIRV translator handles bfloat16 fmul well

[leonling-ll]

In #2951, we change bfloat16 llvm.fmul to float32 llvm.fmul to get the correct result, which is same logic as the elementwise fmul.

The reason could be that the bfloat16 type fmul from triton pipeline is converted to half type fmul after IGC compilation. Examples are uploaded to fmul_bf16.zip.

We should confirm which part leads to this issue. LLVMToSPIRV translator mis-translate bfloat16 type fmul or IGC does not handle bfloat16 type fmul well.

[leonling-ll]

After checking the output of LLVMToSPIRV translator, llvm.fmul %1757, %1831 : bf16 was translated to TypeFloat 21 16 + FMul 21 1377 1239 1314 which is the fp16 type without error reported.

There is no bfloat16 support in SPIR-V for now, SPIRV Translation should raise exception when get bf16 fmul instead of silently accept it.
We should ask LLVMToSPIRV translator team to help with this issue.

[leonling-ll]

Have submitted an issue to SPIRV team.

[leonling-ll]

KhronosGroup/SPIRV-LLVM-Translator#3047 will raise UnsupportedLLVMBFloatType to handle bf16 fmul case.