adding focal loss, distance map dice loss, loading distance maps

[andy-s-ding]

Added a FocalLoss class that also works in compound loss functions. nnUNet also currently does not have a framework for loading in distance maps that were calculated offline. I propose a framework of calculating distance maps from segmentations in the preprocessed folder designated by the plans json. These maps are saved as <case_num>_dist.npy. In doing so, the plans json does not need to necessarily be modified. I have also made a Dataset class and a DataLoader class that is able to load these distance maps during training. Furthermore, I have added custom transform classes that transform distance maps commensurately with image data and segmentations. To make this work, I submitted a pull request in batchgeneratorsv2 to modify the BasicTransform class slightly.

[andy-s-ding]

I also added compute_distance_maps.py in the preprocessing folder. One major TODO is to integrate distance map computation into the preprocessing pipeline. For now, I am taking the _seg.npy files after preprocessing, computing the distance maps, and saving as _dist.npy files.

Currently training distance-weighted loss functions right now and can confirm that epoch time is essentially the same (56.7 seconds per epoch for CE + DistDiceLoss vs. 51.2 seconds per epoch for CE + DiceLoss) on an RTX 3090 GPU.

Again, based on my edits, this is contingent on BasicTransform identifying 'dist_map as a key in the data_dict. I have submitted a pull request for this change in batchgeneratorsv2 (MIC-DKFZ/batchgeneratorsv2#10)