(ignore everything below, it's outdated now).

This is a pretty messy PR so far, but there are a lot of changes I had to make.

Right now, I'm in the middle of reorganizing the plumbing of genv::no_panic so that when it's doing the parent-walk, it returns when it finds either a parent item who's annotated with flux::no_panic, or an FnSig, in which case we can just return the fn_sig.no_panic(). While doing this generalization of no_panic to return an Expr instead of a boolean, I ran into a query cycle where conv_fn_sig is calling genv::no_panic, and vice versa.

So I think I'm violating some scheduling invariant. Maybe we need to move no_panic_if so that it's an attribute again, like with the regular no_panic? I don't remember all the details why the two annotations are handled so differently.

Okay, I think getting a basic version that doesn't die is alright. Some todos to continue with:

• support multi-closure inheritance (see the "grandchild" test of the last commit)
• add the condition Nico told me about the two binders of the FnSig
• write test to see if more complex no_panic_if expressions blow the system up, e.g., no_panic_if(T::foo_no_panic)
• do a merge with the syntactic-inference PR as a stress test. ideally the alarm.rs closure false positive should shut up.