Fix a stack-discipline bug with inlining coroutines

[rjmccall]

Stack-allocating instructions in a function are required to obey a local stack discipline. Inlining a non-coroutine maintains this stack discipline in the caller, but coroutine inlining doesn't necessarily. That's because memory can be allocated after the begin_apply and deallocated after the end_apply, and in fact this is a somewhat likely output of SILGen if a temporary is required during an access. (Memory can also be allocated before the begin_apply and deallocated before the end_apply, and of course we need to handle that, but it's not a likely output of SILGen.)

One potential fix would be to require coroutine begin/end to participate in stack discipline, but that would create some pretty big problems for SILGen (which manages both access lifetime and temporary lifetime but uses different cleanup-stack mechanisms for them) and probably also the optimizers. So instead we just clean up stack lifetime by delaying deallocations using the existing StackNesting framework. As a compile-time optimization, we can avoid this if the callee doesn't have allocations that are live across the yield or if the caller's stack is the same at the end_apply sites as it was at begin_apply.

[rjmccall]

@swift-ci Please smoke test.