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

[via llvm-commits]

vporpo wrote:

I think you are right, we should keep this precondition. Having any other type of instruction in the sequence other than waitcnts would probably make the transformation illegal. 

One change that would help enforce this rule in the future is to make a small change in the structure of the code like this:
```
} else if (Opcode == AMDGPU::S_WAITCNT_VSCNT) {
...
} else {
   llvm_unreachable("Expected a sequence of waitcnts!");
}
```
and a similar change in the other overload of applyPreexistingWaitcnts().


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