ENH: Adapt cerebparc segmentation script for fastigial Hausdorff loss

[jhlegarreta]

Adapt cerebparc segmentation script for fastigial nuclei segmentation using Hausdorff loss.

[jhlegarreta]

A few notes:

• Dice is 0 after 300 epochs (after having varied the learning rate to larger values e.g. 1e-3 since at epoch 200 it was at 0). So nothing is learned, as opposed to PR ENH: Check cerebseg script using Hausdorff loss #38. Not sure if changing the value of the alpha parameter would make a difference.
  Note: remember that I trained with No-QC'ed b100-b2000-b3000. So Ids are different. Need to re-run anyways with all 100 subs.
• So last option is to warm-start with Dice and use try scheduler.