A3C doesn't complete.

When I tried with run_a3c_on_cartpole.sh, I can get training outputs in worker_0/1.stdout. However, the multiprocesses simply never finish in a clean way and there is always one or two python processes hanging there and waiting to be killed by hand.

Start by running run_a3c_on_cartpole.sh and reviewing worker_0/1.stdout alongside the multiprocessing shutdown behavior. Trace which processes remain after training completes; done means the A3C run exits cleanly without Python processes left hanging for manual termination.

Written by the indexing model from the issue text.