[llvm] Reapply "[X86] EltsFromConsecutiveLoads - handle trunc(wideload()) patterns" (#199371) (PR #208999)

[Adam Scott via llvm-commits]

as4230 wrote:

@RKSimon I merged the generic emission into this branch and it folds almost everything as well as before except a handful of 512-bit source stride 8 cases. I closed #208869 since its now just a peephole for those, on shapes only this emits.

Can you take another look at this as-is? Im hoping the generic `trunc(concat(trunc(x), trunc(y)))` emission is the further simplification you were after.

https://github.com/llvm/llvm-project/pull/208999