Fix race condition between handshake and waker registration

[mmastrac]

Ensure that wakers are registered in a thread-safe manner.

This is a latent multi-threaded bug that could happen, though it's a very, very tight race to win.

In a multi-threaded runtime, there was a very small window for a race:

1. The send_handshake part of the handshake task completes, and the handshake then tries to wake any deferred wakers (which do not exist). The task is about to complete but then the thread gets parked by the CPU.
2. In another thread, a poll_read operation (or similar) tests to see if the handshake finished, sees that it isn't and tries to register its waker. The handshake task has already tried to wake the deferred wakers, so nothing ever tries to wake them again.
3. The handshake task completes, but no wakers are woken to complete the handshake process, locking up the connection.

This wasn't easy to repro on my Mac -- I had to explicitly add a thread sleep after the poll_read check for handle.is_finished(), but before we set the waker. This would trigger the race pretty much every time.