Base classes for environments

[MichelangeloConserva]

How should we structure the base classes for the sequential environments?

I don't particularly like the approach in the sequential-neural-testbed repository. They assume that the dataset is fixed and simply iterates over it. I believe that a more flexible approach that allows for the data to be sampled (and maybe batched to improve sampling speed) would be more convenient. In other words, the SequentialEnvironment base class should take a data generator function in input and iteratively sample data from it.

[MichelangeloConserva]

The base.py file contains the current base class for sequential environment.
It is vaguely inspired by the sequential-neural-testbed base class.