Improve robustness of directed message sending

[dforsten]

In the ethcore/sync/src/api.rs file "send()" function we assume there is always a session object for members of the "connected_peers" collection.
This assumption does not hold in practice, and results in crashes at run-time.

Remove the "unwrap()" calls in the following code and replace by proper error reporting:

			let node_ids = peer_ids.iter().map(|&x|context.session_info(x).unwrap().id).collect::<Vec<Option<H512>>>();
			let index = node_ids.into_iter().position(|r| r == node_id).unwrap();

[afck]

The consensus algorithm also relies on all messages between correct nodes eventually getting delivered, so maybe we should cache the messages, and resend them once we have connected to the peer in question. If we fail to do that, we count towards the faulty nodes.

[dforsten]

Good point, that was one of the tasks I prepared Vinu to be coming up.

If a validator node crashes it will lose that cache, but by definition it would then be a faulty node.

[dforsten]

A new issue for caching and re-sending of messages to temporarily disconnected reserved peers had been created and assigned to Vinu here:
#175

[vgtom]

OK so basically in ChainNotify::Send...while doing context.session_info(x)...if i do not get back a valid session, then i cache this message and node_id(PeerId may not be valid at later point) and then send this message at a later point in time when the session is infact established??