[llvm] [AMDGPU] Ensure v_mfma_scale_f32_{16x16x128|32x32x64}_f8f6f4 instructions are convergent (PR #178627)

[Frederik Harwath via llvm-commits]

================

----------------
frederik-h wrote:

I have added tests for other kinds of MFMAs since I did not find a test which already covers the machine-sink/"convergent" behavior. The number of instructions feels a bit overdone. Should I reduce the test?

https://github.com/llvm/llvm-project/pull/178627