[llvm] [AMDGPU] Prefer v_madak_f32 over v_madmk_f32 to reduce vgpr pressure (PR #72506)

[Jay Foad via llvm-commits]

https://github.com/jayfoad commented:

I did some analysis of this across 10000 graphics shaders. The numbers are small but they all seem to move in the right direction:

- Total number of instructions decreased from 7854951 to 7853542 (-0.02%)
- Total number of code bytes decreased from 40989932 to 40981432 (-0.02%)
- Total number of readlane/writelane instructions decreased from 52104 to 51296 (-1.55%)
- Total number of vgprs used decreased from 636518 to 636277 (-0.04%)
- Total number of sgprs used increased from 905361 to 905472 (+0.01%)

https://github.com/llvm/llvm-project/pull/72506