[llvm] [AMDGPU] Test saturated f32 to i8 conversion on vectors (PR #187487)

[Jay Foad via llvm-commits]

jayfoad wrote:

@IgWod would you be interested in improving the codegen for these cases using v_cvt_pk_[iu]16_f32? The potential problems are:

- It's only available on GFX11+.
- There is no "unpacked" scalar scalar form of the instruction (i.e. no "v_cvt_[iu]16_f32"), so we would need some fallback rules to use the packed instruction even for scalar cases.

https://github.com/llvm/llvm-project/pull/187487