Improving (hopefully ;) Serf reconnection ability

Consul agent will try to rejoin only "last seen alive" nodes on restart, which can effectively isolate nodes from the cluster (e.g. see hashicorp/consul#1271) and render cluster inoperable.

I believe this is a limitation of Serf, as Snapshotter contains only alive nodes.

There are several ways to solve this. I can fix it, but I'd like to agree with you on fix design.
For now I'm inclined to:
1. change serf to distinguish "failed" nodes and the nodes that "left" (add failedNodes to the Snapshotter; change Snapshotter.processMemberEvent; change Snapshotter.replay)
2. change serf.Create to rejoin not only previously alive nodes but also nodes that were previously failed

Any comments?