Add utility function to take intersection of features in training and test data

[arnedb]

The number of features returned by a computation graph can be different due to failing transformers (which happens due to errors such as zero divisions). However, machine learning methods expect the same sets of features at training and test time.

Add a function that does the following:

features_train, features_test = intersect(features_train, features_test)

The returned features_train and features_test should have the same column names and order.