Improve auto-scaling machines scheduler
We should improve a little auto-scaling machines scheduler. It should be aware of number of machines in error
state and should delay new machine creations (and remove re-try operations) in such case.
Other way - in case of network problems or cloud provider issues - we can finish in a state where cloud provider limits are used only for machines creation and destruction events and there is no limits left for normal builds processing.
/cc @ayufan