Add fast AUC computation that can be used during training

[kathyxchen]

This will be a small addition to the utilities in Selene. I want to provide a faster AUC computation (compared to sklearn's implementation, which we are currently using) that can run during training. Additionally, I want to add the option to parallelize the performance metrics computations if the model predicts a very large number of classes (over 10^4).

(Have not yet looked into other "fast" computations, e.g. average precision, but these may come up as well.)

[kathyxchen]

def auc_u_test(labels, predictions):
    len_pos = int(np.sum(labels))
    len_neg = len(labels) - len_pos
    rank_sum = np.sum(rankdata(predictions)[labels == 1])
    u_value = rank_sum - (len_pos * (len_pos + 1)) / 2
    auc = u_value / (len_pos * len_neg)
    return auc