Allow re-using the ThreadPoolExecutor across epochs.

[mthrok]

No description provided.

[mthrok]

Update: This turned out to be trickier than I thought. EventLoop shutdown the default thread pool at the end, so either we need to keep the event loop or wrap the executor to make the shutdown request from the event loop no-op.

However, in both cases, we need to think about the cases where the iteration is interrupted in the middle. When there are work items in the queue of ThreadPoolExecutor, and the foreground loop is interrupted, what to do with items in queue?

This corresponds to cancel_future of shutdown method.

https://docs.python.org/3/library/concurrent.futures.html#concurrent.futures.Executor.shutdown