Change failover instance when instance is stopped

Currently, if the instance is stopped, we still check for enough memory
on the target node. This is a little bit too strict, since in case too
many nodes have failed and one is out of the memory, this prevents
fixing the cluster (with the instances down).

We change it to do the memory checks only when the instance will be
started.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>