Allow parallellization over the sliding-windows instead of over the feature functions?

[jvdd]

This is especially of interest for machines with many cores when dealing with a few really expensive feature functions.