Class balancing strategies

Instead of subsampling majority classes, we could weight the different classes so that errors on minority classes are more impactful than errors on majority classes. This would be useful for training single-behavior taggers, such as a Roll tagger, and could make issue #83 (closed) less relevant.

Weighting a class by its normalized inverse frequency would be the new default balancing strategy, as a replacement for subsampling.

The larvatagger.jl train and tagging-backend train commands would admit an optional --balancing-strategy argument with values 'auto' (new default), 'maggotuba' (current subsampling-based default) and 'none'.