Implement failover to other data nodes for distributed queries

[pauldix]

The remote mapper makes requests to data nodes to run their map task. If a data node is down it should switch over to the next one.

The server should also remember which data nodes were recently down and avoid hitting them. At some point they should be back in the rotation to balance load, but it should use some sort of exponential backoff.

The remote mapper request kickoff is here: https://github.com/influxdb/influxdb/blob/distributed-queries/remote_mapper.go#L81

As you can see it has an array of data nodes. There will need to be some way to have that share state somewhere across all the different goroutines/requests.

The data nodes get handed to the remote mapper here: https://github.com/influxdb/influxdb/blob/distributed-queries/tx.go#L157-L163