[RL-baseline] Model v5

[ziritrion]

The policy network for model v5 is completely redesigned. It features 6 convolutional layers rather than 3, and it drops the pooling layers because the convolutions already take care of reducing the dimensionality of the input. It also features 2 fully connected layers right after the convolutions, and both the actor head and critic head feature 2 fully connected layers as well.

This model is based on the model featured in this example . This code is meant for PPO, but since it managed to train successfully without significantly modifying the environment state, we took it as a base to see how it would work for REINFORCE with Baseline.