Add confidence intervals to permutation feature importance

[rogancarr]

Permutation Feature Importance (aka PFI) computes the importance of a feature to a model by permuting values for that feature, scoring it with the model, and comparing the new evaluation metrics to the original evaluation metrics. For speed, PFI uses only one permutation, and this leads to a bit of randomness in the predicted importances. For example, based on the random seed features can change orderings of importance and permutations can even end up showing to improve the model performance. These issues can be fixed by allowing the calculation of confidence intervals around the feature importance values.