Fix server crash detection

[DRMacIver]

Fix the client hanging when the hegel server process exits unexpectedly:

• Drop channel senders in the background reader on pipe close, unblocking any waiting threads
• Add atexit handler to kill the server subprocess on process exit (statics aren't dropped)
• Handle server exit properly in the event loop instead of .expect()

[Liam-DeVoe]

registering that I think our server handling logic in general is dirty and overly complicated and probably missing some cases, but I am happy to go with "it works" for now

[DRMacIver]

registering that I think our server handling logic in general is dirty and overly complicated and probably missing some cases, but I am happy to go with "it works" for now

Yes I agree. I don't like it at all, and I think there are demonstrably bugs lurking - we're getting some intermittent hangs in CI which I'm very suspicious about. I'm hoping that for now we can just cover it in tests until the problems all give up in terror.

Part of the problem is that we've moved to the background thread model for performance and this makes error detection much harder to do. One of the virtues of the cooperative model is that streams just broke under you and you could immediately tell. I think it's worth considering moving back to that but with a better implementation.