[clang] [llvm] [AMDGPU] Add an option to disable unsafe uses of atomic xor (PR #69229)

[Matt Arsenault via cfe-commits]

arsenm wrote:

> @arsenm I agree that the default should be assuming fine-grained is possible. My thinking behind the original naming and direction was not wanting to introduce an unexpected performance regression by default. I'm happy for both to be changed, and this patch being rebased on top of #85052 once it is merged.

The opting for fast-and-maybe-broken by default needs to be a frontend decision (i.e. the language can choose to add the metadata to all atomics). I believe @yxsamliu is going to be working on the frontend half which will preserve the HIP behavior 

> 
> One oustanding question for me is, although outside of the scope of this PR, how will the original 'no-unsafe-fp' option fit in the new metadata node in terms of constraints? Would it imply a new constraint or be covered by `no_fine_grained` and `no_remote`?

I believe we need an additional piece of metadata (which I have another draft for) to express the relaxable floating point. We can express the unsafe-fp-math option with the 2 combined, and then can drop the old IR attribute 

https://github.com/llvm/llvm-project/pull/69229