When autoscaler restarts it has no memory of the previous stats and immediately scales the deployment down. Under my load test conditions,
Autoscaler goes from 7 to 1 then to 2, only then to 6 pods, before finally reaching 7 after a few cycles.
This is obviously not a desired property, so this change fixes that behavior, by
starting Autoscaler in panic mode with panic pod count equal to the current pod count
if we have more than 1 serving pods right now.

Obviously if the deployment is scaled to 0, there's no reason to change logic
and if it is 1, then we won't scale below 1 anyway for the next stable window, so no need to panic either.

/assign @mattmoor @markusthoemmes

This is the GA scope of knative#2930