A couple different, but sort of related changes in this PR:

• On MessagingService shutdown we now close the incoming socket before shutting down the callback expiring map. Although in the previous ordering the callback map won't allow new callbacks to be registered, I think other C* nodes may continue successfully sending messages since at the networking layer they're still accepted. This might help a little bit, in conjunction with the next change, to avoid waiting until timeouts for requests we know will fail -- but probably not a ton.
• If one C* node sees another as up, but cannot connect for whatever reason, it will loop trying to reconnect in OutboundTcpConnection for minRpcTimeout, which can be the same as read/write timeouts in practice. Once this is exceeded, all the messages on the backlog are cleared. If these were needed to satisfy a read/write, this node as the coordinator just waits until the failure callback is invoked via read/write timeout (which can be a long time).
  • If we make this separately configurable and assume recoverable network problems are short-lived, and non-recoverable ones take substantially longer than read/write timeouts, we can make this connectivity timeout small and invoke the failure callbacks much faster than the write timeout, which might improve the client experience.
  • Normally you'd hope this node sees the other as unreachable in this scenario as then it will not attempt to send a request. However, we've seen this not be the case relatively frequently.