[llvm] [AMDGPU] Add new llvm.amdgcn.subgroup.shuffle intrinsic (PR #167372)

[Matt Arsenault via llvm-commits]

arsenm wrote:

I don't particularly like exposing backend intrinsics for high level concepts that don't directly map to hardware concepts. The current main exception to this is for the wave reductions, so this sort of falls in the same bucket. For that case, it's mostly there simply because we needed it for backend internal purposes to hack around DAG issues. If we're going to add these, it should use the wave name instead of subgroup for consistency 

https://github.com/llvm/llvm-project/pull/167372