Handle NaNs

[sjsrey]

This will add functionality to handle missing values encoded as numpy.NaN.

The proposal is to classify the masked values (excluding NaNs), and then set the bin label for the missing observations to -1.

This would keep the labels as ints, rather than using numpy.NaN which would change the labels to floats.

This is a work-in progress PR to solicit comments and ideas.

[knaaptime]

sweet. looks like the only failure is probably small?

[knaaptime]

@sjsrey do you want to get this into meta, or hang on until the next?

[sjsrey]

@sjsrey do you want to get this into meta, or hang on until the next?

Hang on.