Rewrite PCL agent

[lyx-x]

I rewrote the PCL agent to avoid memory issues when saving Variables inside list / replay buffer. I didn't compare the training curve with the old one, but it seems to learn (the average_value increases and R gets bigger) on Catpole under the new parameters and there is no memory issue when run with large network / reasonably long trajectories.

Main methods are the following:

update: take a loss (as an array), log the result as usual and call optimizer (the backprop is done before this function is called)
update_on_policy and update_from_replay: sample a list of trajectories (from replay or the current one), clear grads and compute loss
compute_loss: take a list of trajectories, perform batch computation (batch size is the number of episodes, which may not be efficient when there is one single episode for on-policy update). This function will call backward immediately and only return an array for logging
_compute_path_consistency: compute path consistency, this part of code is almost unchanged

The new underlying data structure is a list of dict to store the current episode, then a replay buffer that only stores (s,a,r) pairs. The old mu (action_distrib) is removed since it can be recomputed again from other items.

I also added a unified model in the example script and changed a couple of parameters.

Issues addressed: #109 #236 #240

I am not sure if the parameters are used correctly, but if they are correct, this PR also addresses #238

[muupan]

Thank you for the improvements on PCL. I haven't checked the implementation details yet, but I think solving the memory issue is great as long as it won't make training slow.

Can you show the training curves and computation speeds before and after this PR?