[llvm] [X86] Try Folding icmp of v8i32 -> fcmp of v8f32 on AVX (PR #82290)

[via llvm-commits]

goldsteinn wrote:

> > @RKSimon, thoughts on dropping this? Do you think its worth the complexity?
> 
> Do you know which part of the fold is causing the most regressions? #82242 was for the limited case of when we've masked the lower bits for comparison - if we just handled that, what does it look like then?

Most of the regressions are caused by not simplifying `CMPP` which #84360 cleans up.
The remainder of when we simplify `ymm` -> `xmm` after the transform which is harder
to undo.

https://github.com/llvm/llvm-project/pull/82290