It is unnecessary because vfd_open will check the handle and return a valid pseudo-fd

for compatibility with stdlib's socket module

for compatibility with stdlib's socket module

add server.DatagramServer; modify baseserver.BaseServer in such a way that makes it a good base class for both StreamServer and DatagramServer

baseserver.py:
- BaseServer no longer accepts 'backlog' parameter. It is now done by StreamServer.
- BaseServer implements start_accepting() and stop_accepting() methods
- BaseServer now implements "termporarily stop accepting" strategy
- BaseServer now has _do_read method which does everything except for actually calling accept()/recvfrom()
- pre_start() method is renamed to init_socket()
- renamed _stopped_event to _stop_event
- 'started' is now a read-only property (which actually reports state of _stop_event)
- post_stop() method is removed
- close() now sets _stop_event(), thus setting 'started' to False, thus causing serve_forever() to exit
- _tcp_listener() function is moved from baseserver.py to server.py
- added 'fatal_errors' class attribute which is a tuple of all errnos that should kill the server

server.py:
- StreamServer: 'ssl_enabled' is now a read-only property

Avoid copying list of links by maintaining a 'dirty' flag that causes
_notify_links to restart iteration. This makes sure that we don't miss
a recently added link because notification is already in progress.

It is possible to notify the same links (that did not remove themselvers)
more than once, but it's better than not notifying some links at all.

rawlink() no longer schedules notification immediatelly if needed, because
there seem to be utility in such behaviour.

That way generic pool methods are grouped together.

without these checks passing invalid arguments would crash the process;
with the checks it raises ValueError.