Connection logic improvements

[horkhe]

This PR brings several improvements and fixes that we made in the connection object implementation to address issues that had been discovered while running the library in production. They are as follows:

• recvTimeout, pingTimeout derive from session timeout and therefore should be recalculated if a new sessionTimeout value is negotiated;
• make logging of the connection life cycle state transitions a bit more verbose to facilitate connection issue debugging;
• watches should be restored on successful authentication only;
• the server connection string reported in connecting event was incorrect;
• session information should not be discarded on connection errors. It is better to spin in the connect/disconnect loop until ZooKeeper cluster recovers then to drop a legit session data.

If you think that these changes should be separated into different commits then let me know. Although some of them just one liners, and the majority of changes were made to write NoQuorum test anyways.

[samuel]

This change was due to some weirdness in ZooKeeper where it would kill the client with no response if an invalid session ID was sent. This caused an infinite loop of reconnects.

I agree that not killing the session on an I/O error is technically correct though.. just depends on if ZooKeeper still exhibits that behavior.

[samuel]

Thanks for the changes. Looks great.. especially having more tests

[horkhe]

@samuel that was fast! Thanks!