Add logic for only computing features from input data

[paulmueller]

The question is how to most-efficiently implement a "copy" segmenter or a submodule that does that more efficiently than just copying the input masks in a Segmenter subclass.