Harvester runs really slow when hit bugs

[jbrown-xentity]

When the harvester hits a hard failure and crashes, it takes ~2.5-3 minutes to build a new machine and get the fetch process started. This use to take ~1 second. This means jobs that have a bunch of failures can tie up the harvester for hours...

How to reproduce

1. Create a harvest source for https://www.usda.gov/sites/default/files/documents/data.json on a cloud.gov machine with harvesters.

Expected behavior

Processes the ~2K datasets in less than 15 minutes

Actual behavior

Still running after 24 hours

Sketch

We could take the "mitigation" approach, and try running the harvest job as a sub-process that gets restarted on failure (using supervisor as we do currently), or just running a bunch of fetch processes.
The long term fix is to make the harvests more robust and just report failures, don't hard fail. This will require much better error handling than is currently implemented, across upstream code and GSA code.