Use Celery Task event methods instead of updating at context manager level

[agjohnson]

Raising this because of #3864 and some issues we hit today.

Currently we're using context managers to try to grab and report all the exceptions that we can. We could actually be using Task.on_failure[0], Task.on_success[1], and Task.on_retry[2] though. Instead of grabbing and reporting exceptions to users through the context manager, instead just bubble these up to Celery and update the build state at task failure/sucess.

This would be tried in steps:

• First, break out the update build code from the context manager. This will live in a utility function so it can be shared. Input is build_pk and the build state
• Next, make an exception subclass that can bubble up to Celery. That is, one that is not reported to users, but raises up through the context managers
• Make VCS locking a bubble up exception subclass, use autoretry_on to retry the task. (Do we mark as build failed here? Possible, but maybe we get away from that pattern)
• Get rid of generic handling of errors outside buildenvs and use on_failure for this

At this point, we should be catching, reporting, and notifiying (we're not, see #3864) through the use of celery.

Next stage, if this is going smoothly, is to reduce the reporting logic on the context managers and see if celery can report these more consistently.

Overall, this will reduce the duplication in logic that we're using to report/update builds, and exceptions will be reported in one spot.

0: http://docs.celeryproject.org/en/latest/userguide/tasks.html#on_failure
1: http://docs.celeryproject.org/en/latest/userguide/tasks.html#on_success
2: http://docs.celeryproject.org/en/latest/userguide/tasks.html#on_retry