Requires constantinpape/torch-em#529

Hi @schilling40 ,
this PR implements the domain adaptation / semi-supervised training functionality.
The main function is flamingo_tools.training.mean_teacher_training.mean_teacher_training, which supports supervised domain adaptation, unsupervised domain adaptation and semi-supervised learning (details are explained in the function signature). The script scripts/training/sgn_domain_adaptation.py shows how to use it for unsupervised domain adaptation.

In addition, I refactored some training functionality, so this also includes some minor changes to scripts/training/train_distance_unet.py. (And I actually uncovered a minor bug along the way, but it doesn't affect prev. trainings). The changes should be fine, but I haven't checked that it actually still works.