[llvm] [AMDGPU] Preserve LDS waits when demoting single-wave workgroup scope to wavefront (PR #207473)

[Austin Kerbow via llvm-commits]

kerbowa wrote:

FYI this should also fix a high priority accuracy issue in Triton, LCOMPILER-2431, which appeared after #187673. In that case the issue is caused by eliminating fences between writes to LDS and TDM on gfx1250. 

https://github.com/llvm/llvm-project/pull/207473