For now at least, these APIs are always used in an asynchronous context,
so stop using the defer mechanism.

Move defer handling out of Broker and into Waker (where it belongs?).
Now the lock must only be taken if Waker was actually woken.

Knocks 400-item run_hostname_100_times from 10.62s to 10.05s (-5.3%).

Now we can run test.sh with MITOGEN_LOG_LEVEL=debug and things just
work.

Mac's SSL seems to have a pthread_atfork handler or similar that does
this for us, no clue if Linux is the same.

Since the Message refactoring from a few weeks back, __iter__ has had
nothing to throw ChannelError if the remote sent _DEAD.

Those fields are always None since the recent fork cleanup work.

14:50:04 E mitogen: mitogen.fork.Stream('fork.7431') crashed
Traceback (most recent call last):
  File "/home/dmw/src/mitogen/mitogen/core.py", line 1287, in _call
    func(self)
  File "/home/dmw/src/mitogen/mitogen/core.py", line 758, in on_receive
    return self.on_disconnect(broker)
  File "/home/dmw/src/mitogen/mitogen/parent.py", line 370, in on_disconnect
    super(Stream, self).on_disconnect(broker)
  File "/home/dmw/src/mitogen/mitogen/core.py", line 721, in on_disconnect
    fire(self, 'disconnect')
  File "/home/dmw/src/mitogen/mitogen/core.py", line 162, in fire
    return [func(*args, **kwargs) for func in signals.get(name, ())]
  File "/home/dmw/src/mitogen/mitogen/core.py", line 1160, in <lambda>
    listen(stream, 'disconnect', lambda: self.on_stream_disconnect(stream))
  File "/home/dmw/src/mitogen/mitogen/core.py", line 1142, in on_stream_disconnect
    for context in self._context_by_id.itervalues():
RuntimeError: dictionary changed size during iteration

Add testimonial from Clojars

-16% reduction in fork cost.

This lets context.shutdown(wait=True) succeed in the master or any
parent.