Now calling tcp_shutdown will in fact close the socket, which cascades into properly cleaning out all the pending requests.
The tcp_server_windows's shutdown logic had to be rewritted (simplified) in order to take this into account.

Conflicts:
	src/python/src/setup.py

If the TCP client connection alarm triggers close to the connection
deadline (but succeeds), the alarm shuts down the endpoint underneath
the existing transport... which is very bad for its health.

If the TCP client connection alarm triggers close to the connection
deadline (but succeeds), the alarm shuts down the endpoint underneath
the existing transport... which is very bad for its health.

Conditionally runs the tests depending on the availability of Python
versions (because Travis is the worst).