[llvm] [AMDGPU] Eliminate unnecessary packing in wider f16 vectors for sdwa/opsel-able instruction (PR #137137)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Also, the attached images in them, also describes the same in generic way, alongwith a solution image based around this patch

This isn't a substitute for MIR tests for this patch 

https://github.com/llvm/llvm-project/pull/137137