Remove code duplication for categorical values to float mapping between DataDescription and automodel.Feature

[lumip]

Concentrating this in DataDescription would be preferred and likely allow quite a bit of simplification of the automodel module.

A problem with this is currently that DataDescription cannot deal with possible missing/NaN values.