It is generally easier to align images that are more similar to each other, so if you intend to keep your analyses separate, then I'd suggest that aligning the images separately may give more accurate results. However, I've never looked into the effects of mixing images from very different populations, so do not have a good intuition about how findings might depend on whether or not you align all scans together or whether you split the subjects into groups. Maybe you could begin by aligning only within group, and if you want to do the regression slope comparison, then you could re-analyse after aligning all subjects together.
Best regards,
-John