More modular approach and some novelties in params and training

[zutotonno]

• Its possible to add validation data and monitor overfitting by means of early stopping, as well as saving the model training at best checkpoints.
• Its possible to launch an extensive grid search, but some more work need to be done.
• Correctly handled the lstm module
• Fixed minor bug during training sampling : if start index + chunk_len > len(file) -> exception
• Allowing different type of training (random and non random sampling) and correctly 1 epoch -> 1 training set exploration.
• Option to add dropout
• Train/Validation evolution saved in .csv, as well as network configuration in json file