[llvm] [NVPTX] Emulate 16-bit floating point atomicrmw when generating PTX (PR #212383)

[via llvm-commits]

modiking wrote:

Change looks reasonable but I don't have much context on the subtleties so deferring to @Artem-B who also reviewed the previous patch.

Also aarch64 CI is failing but that looks like infra so update the branch and see if the issue persists. 

https://github.com/llvm/llvm-project/pull/212383