Alloca lazification pessimization for setjmp is too coarse

Currently it bails if there is a setjmp and any volatile access.

We could try to trace the volatile accesses back to their allocas. This is a risky and weird thing to do. For example, if we fail to trace a volatile access to an alloca, then it doesn't mean that it isn't operating on that alloca. It could be operating on any escaping alloca.