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

[Michael Kruse via llvm-commits]

Meinersbur wrote:

For atomic, it may make sense to annotate accesses are parallel, if atomic relative to code outside the loop (or iterations that are outside the SIMD width). That's the same as for any memory accesses.

For volatile, it could make sense to ignore it. However, I would rather consider such IR malformed. 

>  those accesses will be re-ordered and the number of such accesses may change, which breaks their guarantees.

NOT a guarantee for atomics

https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4455.html

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