Removing improper agent state transitions

[JackUrb]

Overview

Thanks to @jxmsML, I was able to identify a few state transitions that were occurring that shouldn't have been happening. Each were problematic, and in total they meant that human-human chat tasks would freeze up after the first occurrence of two people leaving a chat at roughly the same time (within 2 mins of each other, the average time for MTurk to mark a return).

This fix addresses the two following issues:

• The routing server would sometimes have stale information, and this information was getting sent back to the main mephisto server and putting the wrong state up.
  • Now this is updated to only check for DISCONNECT changes, as this is the only status that the server will ever have that Mephisto needs to read, and anything else causes Mephisto to send a status update to sync the server.
• When a task was returned, the MTurkUnit that marked this return would move directly from ASSIGNED to LAUNCHED, as this is the MTurk HIT status that we use to note a return event in the first place.
  • We now defer the local transition from ASSIGNED back to LAUNCHED, as the Blueprint will decide shortly after the RETURNED status is set whether the Unit needs to be expired. If it does, the Unit will end up in an EXPIRED status, otherwise the Blueprint will call clear_assigned_agent to move the Unit from ASSIGNED back to LAUNCHED

Testing

I was able to reproduce these state issues before this set of changes, but can no longer recreate them on this branch.

[jxmsML]

LGTM! Thanks for fixing this!!! Only one nit to be fixed.