[llvm] [AMDGPU] Fix GFX11 WMMA intrinsic lowering regression for compute kernels (PR #164036)

[Shilei Tian via llvm-commits]

shiltian wrote:

> AMD's ROCm LLVM fork (github.com/ROCm/llvm-project) solved this differently by modifying the pattern classes themselves to accept both bare operands and VOP3PMods-wrapped operands.

Quite surprised that we didn't bring the fix to upstream.

https://github.com/llvm/llvm-project/pull/164036