Investigate using OCL's atomic functions for AtormicCAS and AtomicRMW codegen.

[etiotto]

          > Have we considered using OCL's atomic functions instead of this "manual" approach?

https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic_compare_exchange

https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic_compare_exchangeatomic

These should map to SPIR-V operations and IGC should be handling those.

Thanks @victor-eds for the suggestion. That is something we can take a look at. This PR scope is more narrow and aims to to just remove the redundant barriers we are injecting in the current codegen approach. I will open an issue to track your suggestion though.

Originally posted by @etiotto in #1872 (comment)