When closing pipes, we defer them to be reaped, but also leave
them in the match list where they might be picked up by ep_match,
or leak.  It's best to reap these proactively and ensure that they
are not allowed to life longer once they have errored during the
negotiation phase.

We use overlapped I/O, so we don't need a separate hEvent.

The logic with overlapped structures was fragile as it used
overlapped ios for the connections rather than a single common
one for the listener.  This changes it to be more like POSIX, and
robust against this error.

This seems to alleviate the use after free crashes, although it
does not seem like it should.  Current theory is that this closes
the handle ensuring that it is unregistered from the I/O subsystem,
thus preventing callbacks from firing and referring to objects that
have been freed.