Use neuroreg coreg instead of mri_coreg in hypVINN T2 registration

[m-reuter]

In order to remove the FreeSurfer dependencies in the seg_only path (except for longitudinal, where we still have them), we replace mri_coreg (and mri_vol2vol) calls with neuroreg coreg. This however changes a few things:

• interpolation is tri-linear instead of cubic of the resampled T1
• keep-precision does not exist, and output is converted to float32 always (I think)
• mri_robust_register is dropped as an option. It was basically another implementation of NMI using Powell optimiser and therefore a duplicate.

In the future cubic interpolation will be added to neuroreg and then can be used here. If input precision is important beyond what float32 can do, let me know and that can be changed more easily.
Finally, we will generate quick surfaces based on the registration and then can explore if neuroreg bbreg adds any value here.