[llvm] [AMDGPU] GCNHazardRecognizer: refactor getWaitStatesSince (NFCI) (PR #108347)

[Jay Foad via llvm-commits]

jayfoad wrote:

+1 for the idea of this patch. It will take me a while to review the details.

It would be great to see a test case that demonstrates the shortest-path behaviour. I think this is technically a bug fix, right?

Incidentally there is another form of recursion in `fixLdsBranchVmemWARHazard` where the HazardFn calls back into `getWaitStatesSince`. But fixing that would require a completely different kind of refactor.

https://github.com/llvm/llvm-project/pull/108347