[llvm] [X86][SDAG] Improve the lowering of `s|uitofp i8|i16 to half` (PR #70834)

[via llvm-commits]

qcolombet wrote:

Note: I actually didn't run any test with this change but @dcaballe did (See https://github.com/llvm/llvm-project/issues/67080).
Note2: I have done a quick swipe of the assembly diffs in `llvm/test/CodeGen/X86/fold-int-pow2-with-fmul-or-fdiv.ll` with `llvm-mca` for the sequences that were not immediately obviously better and the perf is supposed to be better. Expert advices are welcomed :)

https://github.com/llvm/llvm-project/pull/70834