Currently Thor jobs spin up an instance (master+slaves) per graph.
This introduces significant latency.
Add an option to allow Thor to wait for a while before spinning down, so that the job can submit another graph to it.
After the timeout, Thor will spin down as normal.