[llvm] [NVPTX] Fix scoped atomic when given runtime values (PR #185883)

[via llvm-commits]

gonzalobg wrote:

Thanks for explaining Joseph.

I think that for the NVPTX target, the following makes sense:

- `syncscope("singleathread")`:
    * operation -> non-atomic operation.
    *  fence -> dropped.
         - Rationale: we don't need to lower `std::atomic_signal_fence` to anything at PTX level.
- `syncscope("cluster")` operation/fence on targets without cluster support: narrow scope to `block`.
    - Rationale: clusters precisely contain 1 block in those targets, therefore `block` suffices.

https://github.com/llvm/llvm-project/pull/185883