test separate-label registration

[jordandekraker]

The proposed method is similar to how ANTs does multilabel registration:

1. split the labelmap into binary images of each label
2. (optional) smooth labels - this helps the registration gradient "catch" which way things need to move if initially they are not overlapping
3. run multicontrast registration, wherein each binary image is a contrast
4. apply the final registration

This is computationally more expensive, but could avoid catastrophic failures (e.g. if label 1 from the moving image is being aligned to label 2 from the fixed image). I would say as long as things are working as intended, there's no need to implement this yet, but I wanted to keep the issue open to remind us to watch for such cases and provide a solution, which has worked well for me in the past.