Rework the Torch model factory

[mmcenta]

Hello,

Yesterday, I noticed that the model factory functions we have for torch can have some problems when trying to use PPO with continuous actions. It should be a quick fix, but we also had these issues with creating neural networks with shared layers - so I think it might be best if we rework how these model factories work. I'll think about it and open a PR later about this, but you can let me know if there are any problems/feature requests for the model factory utils for torch.

[KohlerHECTOR]

This issue is related to @philippe's request to have utilities to manipulate actor, critics, and other networks. In particular to do transfer learning on DQN. So we plan to rework the model factory.
Ideas are welcome in advance of the rlberry hackathon @mmcenta @AleShi94 @omardrwch @riccardodv @brahimdriss @yann .