Output (n_samples, n_classes) for predict_proba

[jwdink]

• Adds supports_predict_proba to Family
• If the family supports predict_proba, then method is added to GLM
• 'Full' classification is only supported if, additionally, the distribution doesn't have a total_count parameter. For distributions like binomial/multinomial, it doesn't make sense to train on / predict classes, because this won't capture heterogeneous counts
• If 'full' classification is supported, you just pass the class-labels and a LabelEncoder does the work. if not (e.g. multinomial) you have to pass a N*num_classes array for y which contains the counts
• For categorical/multinomial, the number of parameters is fewer than the number of classes, since predictions for the final class can be made by leveraging the fact that the sum of the probabilities adds to 1
• Move family aliases from Family to Glm

Resolves https://github.com/strongio/foundry/issues/14

[jwdink]

@andywong36 OK, this was tricky, but I have a 2nd stab at this, take a look at the updated description above, and the updated PR.

[jwdink]

Now that we have is_classifier, I no longer have an excuse not to support R^2 scoring for non-classifiers 😬

[andywong36]

@jwdink I'd like to hold off on this review until I complete #16 so I can better understand what's going on.