Driver node events can be delivered out of order

[talex5]

I added some logging to the overlay driver to log NodeDiscovery events.
Then I killed dockerd (kill -9) on host 172.31.26.11 and restarted it. Here are the log events printed on another node, watching the restart:

Aug 21 07:47:21 ip-172-31-0-214.us-east-2.compute.internal dockerd[10382]: time="2018-08-21T07:47:21.292565622-07:00" level=warning msg="overlay: New/NodeDiscovery(discoverapi.NodeDiscoveryData{Address:\"172.31.26.11\", BindAddress:\"\", Self:false})"
Aug 21 07:47:34 ip-172-31-0-214.us-east-2.compute.internal dockerd[10382]: time="2018-08-21T07:47:34.175288657-07:00" level=warning msg="overlay: Delete/NodeDiscovery(discoverapi.NodeDiscoveryData{Address:\"172.31.26.11\", BindAddress:\"\", Self:false})"

Note that:

1. It gets the notification of the node rejoining first, then the notification that it left second.
2. The notification only contains the IP address.

The overlay driver might not care about this (DiscoverDelete just returns nil), but it might make sense to change this in some way. One option would be to notify drivers about changes in networkdb's view of membership rather than memberlist's view (networkdb removes the old node as soon as it hears about the old one - but see #2256). Another would be to include the node name in the message, so it's clear they are about different instances.