dwi2response: -shells affects brain masking

[Lestropie]

Discovered by change while checking feasibility of [comment in #2960] (#2960 (comment)), due to dwishellmath warning added in #2288 and dwi2mask changes in #2197. Should not be affecting 3.0.x.

The dwi2response interface command applies shell selection during DWI data import, and then uses that image for brain mask derivation if necessary. This means that for algorithms like dwi2response tournier and dwi2response tax, which explicitly strip from the imported DWI data all shells that will not be included in the exemplar voxel selection / response function, that stripped DWI data is also being used for brain mask estimation. This could be particularly problematic if, for instance, a user were to nominate some preferred default brain masking algorithm that is explicitly designed to work on b=0 images.

So either:

• The DWI import should pull data from all shells, and shell selection should either be done as a second explicit extraction call or the downstream response function estimation algorithms should use the manual shell selection wherever relevant;

• The dwi2mask call should be done based on the input DWI series, not the DWI data in the scratch directory that has already removed data from shells outside of the user's selection for response function estimation.