Draft: Fix class imbalance in CE loss

[PatrickRMiles]

This PR adds class weights to our CE loss calculation to counteract the significant overrepresentation of empty/background voxels in each sample.

Needs:

• If we get unlucky and choose a subset of samples which are missing one or more classes, we should not end up downweighting the missing class(es). Every non-background class should probably get the same final weight.
• Testing to ensure this has the intended effect of speeding up convergence