[llvm] [AMDGPU] fix asyncmark soft waitcnt bug (PR #184851)

[Ryan Mitchell via llvm-commits]

RyanRio wrote:

The assertion guarantees it's a safe assumption, but if you're asking if I think its the correct choice going forward, I still think yes. Perhaps the naming could use some improvement as some other things are getting lumped in there, but they still make logical sense. Changing the assertion and adding some catch all case just makes the flow more convoluted.

https://github.com/llvm/llvm-project/pull/184851