[bg] Catch exceptions in workers

[stephenroller]

Patch description
If exceptions happen in (all) background workers, currently training just hangs indefinitely while we wait for a batch from the queue. This patch adds a timeout, and checks background workers for exceptions when that timeout is hit. I had previously tested checking on these exceptions before pulling from the queue, but it was devastating to speed. This patch seems to strike a good balance.

In manual testing, this patch displayed these properties:

• If some, but not all workers threw exceptions, then training continued with just fewer workers. Sometimes there was a print of an exception, but not always. Speed was similar to launching with an equivalent smaller --num-workers.
• If all workers threw exceptions, then there was a LOT of log spew and the program ended. But the log spew did contain the root exception in it and thus a hint for the developer.
• If no workers threw exceptions, speed was the approximately same as before this patch (within random variation on an 8 gpu 8 worker 10 minute run)

Testing steps
Considerable manual testing.

[EricMichaelSmith]

Makes sense, great!