[llvm] [AMDGPU] Replace zext pattern from reg_sequence to cvt_u32_u16 (PR #208045)

[Joe Nash via llvm-commits]

Sisyph wrote:

> > LGTM on the grounds it is a temporary workaround for a compile time issue and we will reintroduce the patterns when the root cause can be fixed
> 
> Thanks! For this patch I would consider it not a workaround though since using cvt_u32_u16 also provide better code quality. Marked a few in the lit tests

My concern with this patch is not code quality regressions, but rather that we should be able to handle v_mov_b16 better throughout the compiler. This patch hides those type of issues, so to investigate them one would have to revert it.

https://github.com/llvm/llvm-project/pull/208045