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

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

I'd really like to see the codegen diff for a benchmark that shows a real perf differences - my gut feeling is something else is being combined as a sideeffect and its nothing to do with the GFNI instruction by itself.

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