Relax the constraint that a single light source has the same center whether it is a star or a galaxy

[jeff-regier]

Suggested in #480. Originally I constrained each light source to have the same center whether it's a star or a galaxy because otherwise a single light source might end up explaining bright spots in totally different parts of the image. But now that's not an issue: we're constraining light sources to remain within a pixel or so of where they are initialized, and I think that's the right way to do it. Relaxing this constraint should reduce model misfit, and @rgiordan suspects it will reduce the number of Newton steps.