Implement losses and metrics for multi-class segmentation tasks

[josafatburmeister]

• adapt model architectures to use a softmax activation layer for single-label and a sigmoid activation layer for multi-label segmentation tasks
• adapt loss functions to work with single-label and multi-label segmentation tasks
• adapt metrics to work with single-label and multi-label segmentation tasks
• exclude padding from metric and loss calculations
• implement normalised surface distance (NSD)