Downscaling

[benbjohnson]

Currently, the autoscaler only scales up and relies on the target application shutting down when it doesn't have work to perform. This works when there is low job parallelism within a single worker (e.g. a worker can handle 1 or 2 jobs at a time) but doesn't work well as it scales (e.g. a worker can handle tens of jobs at a time).

The reconciler should probably accept a min_machine_count and a max_machine_count expression instead of just the max (via expr).