[llvm] [Metadata] Respect volatile & atomic constraints for parallel_access MD. (PR #208839)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

> Shoudn't it rather be Clang to not annotate such memory accesses?

That would also be an option; I guess then we should clarify the wording to explicitly allow this. But I am not sure if that behavior is actually desirable/useful? 

Also, other frontends may run into similar issues, which is why my preference would be to adjust the metadata.

Independently, clang could warn if trying to use with a loop that has atomic accesses

https://github.com/llvm/llvm-project/pull/208839