Generalize masking in registration framework into weighting

[gdevenyi]

Description

Currently, the ITK registration frameworks support binary masks which include/exclude voxels in the calculation of similarity metrics.

This puts a particularly strong focus on getting mask generation maximally correct as over/under segmentation of masks can cause various sorts of misalignment.

I propose a generalization of the masking framework into a "weighting" framework, where weighting images with floating-point values weight how much a given similarity metric includes voxels in their final value. This would allow such things as "fuzzy" edges on masks with uncertainty, and likely other applications unknown to this author.

As there are probably some important technical details and complications I'm unaware of inside the ITK implementation, I hope I can spur a discussion here about possibly implementing this feature.

Additional Information

Upstreamed from a discussion in ANTs, ANTsX/ANTs#1437

[gdevenyi]

The new https://git.fmrib.ox.ac.uk/flange/mmorf_beta framework from FSL takees advantage of this feature.