[llvm] [AMDGPU] Enable reordering of VMEM loads during clustering (PR #107986)

[Carl Ritson via llvm-commits]

perlfu wrote:

> > Reordering SMEM/LDS access is not beneficial as these must always be waitcnt 0.
> 
> SMEM is out-of-order but LDS is in-order.

True. I have rebased, addressed some comments and cleaned up a bit.
I have also enabled LDS reordering.

I hadn't pushed hard on this as I am slightly wary of regresses in some cases.
Unlike  #107986 which I think is basically free.


https://github.com/llvm/llvm-project/pull/107986