Rely on API error messages instead of checking #5
Description
Instead of checking for available memory and jobs in advance, the orchestrator should always try to start a run, catch possible errors, and maybe wait for a cooldown period when a run fails to start.
This behavior would result in less API calls.