Tensor interpolation

[cookpa]

Came up in the discussion of #24

It's generally accepted that it's good to linearly interpolate tensor components in the log space, but this leads to very annoying interpolation artifacts along the edge of the brain mask.

In ANTs, this is handled by allowing the user to set a "background MD", and background tensors (all components exactly zero) are set to an isotropic tensor with the specified MD.

Another option would be to use nearest neighbor on the boundaries. Given the moving mask, one could compute both linear and NN resampling, then substitute NN interpolation in voxels on the edge.

It doesn't look like there's a smarter way to do this in ITK

https://discourse.itk.org/t/resampling-mask-to-avoid-0s/7304

[stnava]

hi @cookpa

here is an attempt at implementation your suggestion -- or part of it

e3ea114