Class distribution after partitioning

[Yang-Li86]

When a detection dataset contains quite a few classes, for example, 12 classes in totD, the partitioning script inevitably creates clients where for a certain class, there are no bounding boxes at all.

This creates issues with the test.py as there are missing classes, and the class names in the generated CSV won't match the real results. I need to modify it so it obtains class names from the results if I can. However, the mean results (name: all) remain accurate.

[aymuos15]

Basically need a very nice (and easy) way to stratify things.