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

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

... It's not that target specific? I'm pretty sure some flavor of SDWA/Opsel is available on basically everything gfx9+ if not earlier? And also, the DAG legalizer etc. already have a lot of target-specific code in them?

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