Rolling back from a broken update with only one replica fails with a
timeout in the existing code.
The problem is the scale down logic does not consider unavailable
replicas in the old replication controller when calculating how much to
scale down by. This leads to an obvious problem with a single replica
when min unavailable is 1.
The fix is to allow scaling down all unavailable replicas in the old
controller, while still maintaining the min unavailable invariant.
e5a558c35f