Skydns is not resilient for cluster changes

[ederuiter]

We have a setup where skydns only connects to the local etcd2 instance, in this case every cluster error directly affects the client (as the etcd client won't do automatic failover/retry). In order to make this more resilient, it would be enough to call Sync() on the etcd client to get all endpoints from the cluster (even better is to call AutoSync). This way the etcd client automatically does a failover to another endpoint in case of loss of connectivity.

I have a basic implementation in the commit below:
ederuiter@6da0690

Note: in case of writes to etcd, the write can even fail if the local etcd2 instance is still running, but the leader is stopped/rebooted/not-available. So without a call to Sync the write would fail during the leader selection, with the call to Sync the write would be automatically retried.