When performing an info query via request(), a listener is started, and
a packet is formed. As the packet is formed, known answers are taken
from the cache and placed into the packet.  Then the packet is sent.
The packet is self received (via multicast loopback, I assume).  At that
point the listener is fired and the answers in the packet are propagated
back to the object that started the request.  This is a really long way
around the barn.

The PR queries the cache directly in request() and then calls
update_record().  If all of the information is in the cache, then no
packet is formed or sent or received.  This approach was taken because,
for whatever reason, the reception of the packets on windows via the
loopback was proving to be unreliable.  The method has the side benefit
of being a whole lot faster.

This PR also incorporates the joins() from PR python-zeroconf#30.  In addition it moves
the two joins() in close() to their own thread because they can take
quite a while to execute.

This fixes a race condition in which the receive engine was waiting
against its condition variable under a different lock than the one it
used to determine if it needed to wait.  This was causing the code to
sometimes take 5 seconds to do anything useful.

When fixing the race condition, decided to also fix the other
correctness issues in the loop which was likely causing the errors that
led to the inclusion of the 'except Exception' catch all.  This in turn
allowed the use of EBADF error due to closing the socket during exit to
be used to get out of the select in a timely manner.

Finally, this allowed reorganizing the shutdown code to shutdown from
the front to the back.  That is to say, shutdown the recv socket first,
which then allows a clean join with the engine thread.  After the engine
thread exits most everything else is inert as all callbacks have been
unwound.

With the restructure of shutdown, Listener() now needs to throw EBADF on
a closed socket to allow a timely and graceful shutdown.

Also adds names to the various threads to make debugging easier.

Add more needed shutdown cleanup found via additional test coverage.

Force timeout calculation from milli to seconds to use floating point.