optimize speed with labelmap resolution

[jordandekraker]

In the current code, I believe synthseg inference resamples to 1mm isotropic, then resamples back to the native resolution when generating a labelmap. Instead, we could keep these at 1mm for faster registration (since upsampling the labelmaps shouldn't really make a difference anyway).

steps would be:

1. resample moving and fixed to 1mm
2. synthseg
3. label-label registration
4. image-image "2 step" or "fine-tuning" registration at original resolution using the initial low-res label-label registration (SyN only)