[llvm] AMDGPU: Undo atomicrmw add/sub/xor 0 -> atomicrmw or canonicalization (PR #87533)

[via llvm-commits]

b-sumner wrote:

> > > I believe scope does not necessarily point to a specific bus. Most of the uses probably would not request a device scope on an atomic in remote memory, but this is not necessary. Why not to do it for any scope?
> > 
> > 
> > Because the question of bus is only relevant at system scope. The device local scopes don't have the issue
> 
> Adding Brian to double check on this, in the meantime LGTM.

The scope is only a hint about where the target memory actually lives.  But I am OK with the change as-is until there is a complaint.

https://github.com/llvm/llvm-project/pull/87533