[clang] [Clang][AMDGPU] Stop defaulting to `one-as` for all atomic scopes (PR #120095)

[Pierre van Houtryve via cfe-commits]

Pierre-vh wrote:

> > Ah, that makes sense. It's surprising it wasn't like that in the first place
> 
> But can we now get rid of the -one-as sync scopes and replace them with MMRAs?

MMRAs are droppable so not currently suited to replace syncscopes like that, but it's a goal.

As part of the Vulkan memory model work I need to reevaluate that aspect of MMRAs and fix it (= promote them to an instruction operand), or use another system altogether (in which case MMRAs may just get removed). It's not yet decided which one of those 2 options it'll be yet, but I plan to regularly follow-up on this from now on until it's addressed so hopefully something will happen medium term.

I'm not sure how much resistance there'd be upstream to making MMRAs an instruction operand instead of MD?

https://github.com/llvm/llvm-project/pull/120095