[llvm] Use 256 bit register when AVX2 or higher is available. (PR #91721)

[via llvm-commits]

shamithoke wrote:

@RKSimon 
Let me find a stable bench-marking worthy machine and rerun test. If I can reproduce the improvement, I will move the code to either X86FixupInstTuning or MCCombiner, as you recommended. I will close this PR. 

And you make a fair point about the possibility of scheduler models not being updated with GFNI. 

https://github.com/llvm/llvm-project/pull/91721