Fix: signal handler musn't depend on the event loop

[ysbaddaden]

Only a limited set of POSIX functions are signal safe, and the system functions that the event loop implementations can rely on isn't in the list (e.g. epoll, kevent, malloc, ...).

Now, the writer side of the pipe is blocking, so we should never reach a nonblocking case that would trigger an event loop wait, but going to the event loop may still be doing far too much or dangerous things: an event loop might not be available (e.g. bare thread) and it might be lazily allocated (signal unsafe).

[ysbaddaden]

Aside: it's becoming more and more evident that we should move on to signalfd and EVFILT_SIGNAL, and overall to move all the signal logic into the Crystal::EventLoop API, so they can each use what would be best for them.