Hard-Coded and Unnecessary segments should be removed

[UtkarshMishra04]

Current Setup:
There is a lot of hard coding which brings a lot of problems when trying to generalize the framework across multiple settings.

To be introduced:
Code will be cleaned and hardcoded segments will be removed, proper parameterization in Configs will be used for all the variables and hyperparameters.

Major Changes:
The overall task of running custom settings with the default config template will be much easy after this.

Eventually, this will be progressed with #157 and will also resolve #43

[sergiopaniego]

I believe we should deeply review the RL brains removing any unnecessary code and generalising them to easily train-test the agents. What do you think? It's important to already consider this issue for #157! We could add a table in the documentation (Github pages webpage) with the information about available brains and how to train-test them. I thinks it's kind of difficult to understand the RL brains now. Tell me what do you think!

[UtkarshMishra04]

I agree with your suggestion.
Now with more and more algorithms being introduced, the number of hyper-parameters increases substantially, so it becomes necessary to remove any hard-coded segment.
Further, with the documentation, things can be easily tested, verified, and reproduced.

Just a concern, the DQN implementation by @dcharrezt is not applicable with the current settings as it is in Keras. So I propose we completely shift to pytorch as that will resolve the Keras issue and also ease up the CUDA and Tensorborad integration.

I will start working on this, just some minor updates to #157 are pending on my end.

[sergiopaniego]

With this changes that you suggest, the DQN implementation would be just in PyTorch?
We need to still support Tensorflow-Keras for the non-RL part since all the brains use it.