[llvm] ORC: Fix flaky OrcLazy tests (PR #228619)

[Lang Hames via llvm-commits]

https://github.com/lhames approved this pull request.

Short version: LGTM.

Longer version: I'm pretty sure this fix works, but if I'm reading it right the logic only applies to `CV`: The waiting thread has to acquire the mutex and can't do that until the last callback releases it. Since callbacks don't access the mutex a second time the mutex should be safe. That doesn't apply to CV though: there's nothing to stop the waiting thread from tearing that down if the last callback gets descheduled right after it releases the mutex, and right before it calls `CV.notify_one();`.

https://github.com/llvm/llvm-project/pull/228619