Support signal-based orthogonalization with voxel-wise regressors

[tsalo]

Summary

I didn't include this in #1255 because it was going to require a bit of effort and I didn't foresee many people using these two elements together, but I do want to support it at some point.

We need to support the following cases:

1. A voxel-wise regressor is a signal regressor, so we need to orthogonalize tabular regressors with respect to the voxel-wise regressor.
   • In this case the voxel-wise regressor won't be modified, but it will be dropped from the list of confounds files and the design matrix, since we don't pass signal regressors out of GenerateConfounds.
   • All of the tabular noise regressors have to be changed into voxel-wise regressors here too. That's going to be weird.
2. A voxel-wise regressor is noise, so we need to orthogonalize it with respect to tabular signal regressors.
   • In this case we just need to modify the voxel-wise regressor and save it out to a new file.
3. Combinations of the above two. Maybe you have 4 regressors: 1 noise voxel-wise, 1 noise tabular, 1 signal voxel-wise, and 1 signal tabular.

Additional details

I can't think of any actual use-cases for this, but I'm sure someone will at some point.