Allow unconditional failovers off offline nodes

Allow unconditional failovers off offline nodes

Normally, we should not place instances on nodes that do
not have enough disks. However, there is one exception: if
we failover an instance from an offline node, that node can
well be secondary of that instance---the fact that it is currently
primary proves it has enough disks space. The reason why we have
to handle that case special is that if a node is offline, we sometimes
cannot determine the amount of disk available, hence the conservative
estimation is 0.

Signed-off-by: Klaus Aehlig <aehlig@google.com>
Reviewed-by: Hrvoje Ribicic <riba@google.com>