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

[via llvm-commits]

vporpo wrote:

> The crash is a byproduct of the fundamental issue

The code in applyPreexistingWaitcnt() assumes that the only instructions between OldWaitcntInstr and It are waitcnts, which doesn't look like a very safe assumption to make. So I am wondering whether removing this assumption is the better fix. Wdyt ?

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