Adds redis.selector, a module that provides the best selector strategy
available on the current platform. A redis.selector polls a socket to
provide two pieces of functionality:

1. Check whether data can be read from the socket. Prior versions of redis-py
provided this behavior with just select.select(). select() has lots of
limitations, most notably a limit of ~1024 file descriptors. Now that
better selectors are available, this should make can_read() faster and
able to accomodate more clients. See #1115 and #486

2. Check whether a socket is ready for a command to be sent. This doubles
as a health check. It ensures that the socket is available for writing,
has no data to read and has no known errors. Anytime a socket is
disconnected or hung up, data is available to be read, typically zero bytes.

ConnectionPool.get_connection has been modified to ensure that connections
it returns are connected and are ready for a command to be sent. If
get_connection encounters a case where a socket isn't ready for a command
the connection is reconnected and checked again.

TODO: more tests for this stuff. implement EPoll and KQueue selectors.

Fixes #1115
Fixes #486

Do not leave connections in invalid state

Use IPPROTO_TCP constant instead of SOL_TCP constant

Fix github pull template typo

Since Connection.disconnect() now verifies that the current process owns
the connection before shutting the socket down we can safely readd
the destructor just to make sure things are really cleaned up