background mask option for pseudo labeler - unsupervised training

[stmartineau99]

• added new class NewPseudoLabeler to subtract background from the pseudo labels if a background mask is given
• updated get_unsupervised_loader to handle 4 cases:

1. both sample mask and background mask are given, this will be be used by the train loader
2. sample mask only, this will be used by the validation loader
3. background mask only, not implemented error
4. neither mask given, use the default behavior

• created new classes to ensure that get_unsupervised_loader handles each case correctly: DropChannel, ChannelWiseRawTransform, ChannelWiseAugmentations
• updated mean_teacher_adaptation to use the NewPseudoLabeler if the background mask is given
• added subclass NewMeanTeacherTrainer, which drops the background mask from the teacher input after computing pseudo labels. it also drops the background mask from the student input since this is not used. This behavior only occurs when it recieves training data, and not for validation.
• fixed the NewPseudoLabeler so that it behaves correctly when it recieves training and validation data, since in the case of validation there is no background mask channel.

[constantinpape]

This looks very good now! The only things to address:

• You are not actually using the new class for the mean teacher trainer.
• Some cosmetics.

See details in the comments.