Make RPC handling more robust while rolling Consul servers

[slackpad]

When rolling servers there are many ways that applications can see 500 errors. There are mechanisms like the RPC hold timeout, but there are some other edge cases we can improve. This issue is still being defined, but is open to capture areas we can improve.

• Have the server hang around for a configurable little while after the leave completes to let existing RPCs finish out. We could also have servers send an event that shifts traffic away before they even attempt to leave, which would be even better since the leave already takes a few seconds to complete.
• Clients use the server manager and will shift away from a server when it leaves, but servers use the Raft information to do this. The previous item should cover the server case as well, though we may want to tie in the Serf data to more quickly send "no leader" errors (and thus retry) if Raft thinks there's a leader but Serf thinks it has left.
• The connection pool EOF case added in https://github.com/hashicorp/consul/blob/v1.0.0-beta1/agent/pool/pool.go#L417-L423 should also apply to the error case in https://github.com/hashicorp/consul/blob/v1.0.0-beta1/agent/pool/pool.go#L408-L410.
• We've seen cases when the leader goes down where you get an EOF error at https://github.com/hashicorp/consul/blob/v1.0.0-beta1/agent/pool/pool.go#L406-L410. This is a safe place to retry for another server since the RPC has not yet been made. Adding a retry here in addition to the above could make things way more robust.