Don't rely on an internal flag

[takluyver]

This fixes #2 (I hope :-), a race condition between changing the flag value and updating the operating system eventfd/pipe/socket. Rather than trying to always update them together using a lock, this drops the flag, and uses the operating system whenever we want to query or update the event state.

This also means that an EventFD object can be used reliably between processes - child processes will share the operating system state but they would each have their own _flag value.

Non-blocking FDs are used so it doesn't hang if multiple threads/processes try to set or clear the same event.

[takluyver]

The remaining failures are all on Python 2.7 because it doesn't have the selectors module (and almost certainly other issues). It's all fixable if you really want, but there are a lot of little niceties in recent versions of Python, and we're a year past the end of life for 2.7 now.