[clang] [Clang][AMDGPU] Add -f[no-]atomic-backward-compatible and refine atomic codegen (PR #157672)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

Closing this for now. If a concrete need comes up later, we can revisit the issue and choose a simpler approach based on the current behavior.

https://github.com/llvm/llvm-project/pull/157672