Fixed very vexing bug that would sometimes cause a slice to be labelled as outlier in all volumes in a shell. Not fully tested yet.

Fixed very vexing bug that would sometimes cause a slice to be labelled as outlier in all volumes in a shell. Not fully tested yet. Sometimes a slice was labelled as outliers because the prediction had "too large" values (caused by large Jacobian determinants) in a set of voxels. Once that had happened the predictions would now be biased towards too large values in those voxels, which in turn could cause that slice to be labelled as an outlier in some mode volumes, etc etc. Eventually, after enough iterations, that slice could be labelled as an outlier in all volumes. That would in turn mean that no shape-reference volume could be found (as reported many times by the HCP people) and/or very large intensity values in the eddy corrected data in those voxels (usually small clusters of voxels on the edges of the ear-canals). It has been solved by adding a mask to the comparison between observed and predicted data. The mask is based on the Jacobian determinants, so that the comparison is only based on voxels with Jacobians greater than zero and below some threshold (default 3.0).