Split compartment postprocessing into separate tasks

The prediction model always generates both compartments, and puts them into a single file (2 pairs of predictions, 4 arrays total). Then, postprocessing on each is independent.

Instead of saving to a shared raw predictions file, and later to a shared predictions file, we should separate everything into the compartment. For example,

raw-predictions-whole-cell.npz
predictions-whole-cell-mask.npz
predictions-whole-cell-shapes.json

This would let computation happen in parallel, and also reduce memory requirements by ~2x pixel size.