Close environment if step raises an exception.

[chriselion]

In #2988, that was raising with this callstack:

Traceback (most recent call last):
  File "/Users/chris.elion/code/ml-agents/venv/bin/mlagents-learn", line 11, in <module>
    load_entry_point('mlagents', 'console_scripts', 'mlagents-learn')()
  File "/Users/chris.elion/code/ml-agents/ml-agents/mlagents/trainers/learn.py", line 422, in main
    run_training(0, run_seed, options, Queue())
  File "/Users/chris.elion/code/ml-agents/ml-agents/mlagents/trainers/learn.py", line 266, in run_training
    tc.start_learning(env)
  File "/Users/chris.elion/code/ml-agents/ml-agents/mlagents/trainers/trainer_controller.py", line 209, in start_learning
    n_steps = self.advance(env_manager)
  File "/Users/chris.elion/code/ml-agents/ml-agents-envs/mlagents/envs/timers.py", line 262, in wrapped
    return func(*args, **kwargs)
  File "/Users/chris.elion/code/ml-agents/ml-agents/mlagents/trainers/trainer_controller.py", line 283, in advance
    step_info.brain_name_to_action_info[brain_name].outputs,
  File "/Users/chris.elion/code/ml-agents/ml-agents/mlagents/trainers/rl_trainer.py", line 131, in add_experiences
    curr_to_use, take_action_outputs["action"], next_info
TypeError: 'NoneType' object is not subscriptable

The actual error isn't that important, but the thing to note is that the worker process stuck was waiting for the main process :

ml-agents/ml-agents-envs/mlagents/envs/subprocess_env_manager.py

Line 104 in 2a0bb71

cmd: EnvironmentCommand = parent_conn.recv()

(I think, may have been somewhere different).
This means that the process never exited and basically deadlocked.

This change tries to call EnvManager.close() even if training raised an exception. I confirmed by undoing the fix from #2988 that this at least lets python exit and Unity stops playing (instead of having a spinning beach ball).

[chriselion]

close() isn't being called here anymore. I think it's cleaner to have it in the trainer_controller since that's what creates it in the first place.

[harperj]

👍 thanks for this fix, looks good!