Refactor CLI options for nnunet

[hermancollin]

For the nnunet migration, some CLI flags will become useless.

• -t is already almost useless and I don't remember why we still require it;
• -s will not be necessary anymore because we do not resample the input images anymore. Although we might want to keep the -z zoom option to still support input resampling. It will however require a new implementation because it currently depends on the ivadomed inference function;
• --overlap might not be necessary anymore (I don't think it's widely used anyway)
• --no-patch will not be relevant anymore. Haven't seen the kind of patch stitching artifact we had with ivadomed so I don't think it would be useful.

There might potentially be new options, though. We could have one to specify which checkpoint to use (best or final, see #797), but I would prefer not to expose this to the user such that we keep total control of what models are used. I don't think they need to know about these details in the first place.

[hermancollin]

As of today, the -t, -s, -z, -r, -l, --no-patch and --overlap options are no longer available (see #800). This simplifies the segmentation CLI: the following command is now valid:

axondeepseg -i input.png

Looks much cleaner! Note that this was one of the main motivation to train a single model.

[hermancollin]

We could also remove the -i option since its the only required arg.