Feed data to `train` instead of `ESN`

[MartinuzziFrancesco]

Pretty much title, the goal is to divide the definition of the model and the actual training. We could also add an instantiate function to run the data through the ESN to obtain the states without the training