Provide autoscaling of nodes

[mschubert]

The worker pool does not have to be fixed size, but could be extended if all nodes are saturated or decreased if they are not.

This would likely require:

• The master to track how many workers are submitted vs. are taking calls (we already do this)
• The master to check the load of workers during evaluation (refactoring the shutdown stats into a callable)
• The workers (QSys class) to be able to add/remove nodes (basically, just call submit or WORKER_STOP when required)

Related: https://cloud.google.com/kubernetes-engine/docs/concepts/cluster-autoscaler