Add composition to validity metric

[sgbaird]

E.g. distance calculation based on fingerprint + threshold

Could be a scalar featurizer (seems most straightforward, use EMD between distributions, normalize by some dummy metric, and average the two distributional distances).

Could be smact_validity, but probably need a way to do a logical AND match (i.e. change structure validity to be a YES/NO on a per-structure basis - also, it's no longer comparing to train+test at that point, it's standalone).

#44