Fix ML processing that is causing jobs to fail for medium-large jobs

[mihow]

A continuation of #773

Even after switching to RabbitMQ, our processing jobs are still failing after 100 images or so.

I believe this is because of the huge responses introduced in #684. They are collected into results_json in the save_results_async tasks and appear to reach over 30mb each.

However, now we have also observed this behavior for some large export jobs, which don't have huge result responses (everything is written to a file). More investigation is needed.

[mihow]

Some notes and things to try after discussion with Francis

• Switch from request & response method to queue that worker can pick from

• Switch from one long running master task with sub-tasks to many tasks (one per 4 images, etc) and use Queues

• Save results to a file for Antenna to ingest, rather than returning in response

• Are workers crashing? and then not telling the master task that they are no longer processing that task

• Put ml backend and celery worker on same server and communicate with localhost