The `Family` class, decentralised random states, a better caching framework

[daffidwilde]

This is the first big improvement in a while and makes the library easier to use in my opinion. The changes can be summarised as follows:

• A new class Family for handling the subtypes of a Distribution class externally. Note the move from edo.families to edo.distributions to reflect this and to avoid confusion in the future.
• The Individual class is more robust now, taking its fitness as an attribute and being able to store and recover itself entirely -- including the subtype instances in Individual.metadata (a big improvement over the dictionaries).
• The pseudo-random number generator framework has been totally decentralised. This means that stochastic fitness functions can use np.random.seed if necessary without throwing reproducibility out of the window. Individual and Family instances are provided their own np.random.RandomState instances to use in sampling, and there is a "lead" state used by DataOptimiser.

[daffidwilde]

All looks good now the builds are going through. Just need to go over the docstrings before merging.