Scale up/down GPU instances based on n_jobs

[lefnire]

Currently only spins up 1 AWS Batch instance when users are active (spins down after inactivity). Need to scale this.

• Submit additional Batch jobs based on n_new_jobs vs n_machines in ("pending", "on"). Each instance can handle 2 jobs at a time (will improve after Reconsider current BERT models #28), so something like if n_new_jobs/n_machines/2 > 1: cloud_up().
• Inverse check to spin down instances. Move from instance handling its own death, to submitting a kill job from server_jobs (since only one instance will take that job, and server_jobs will re-consider submitting another kill-job after one goes down).
• (small) if 0 in queue, it says "eta 0 seconds" - should be total+30
• Add dict (map) of number of concurrency specific jobs can handle. Eg, summarization/sentiment can handle 2-3 jobs at once; question-answering only one (can't run concurrent with anything else). Also add config.yml option for altering this per machine, in case some machines have higher/lower GPU compute.