datasets.py standardization

[PavlosMelissinos]

1. Better couple properties of the dataset together. Conversions between IDS <-> CIDS <-> CATEGORIES <-> PALETTE should be done in a different way.

2. The only difference between MSCOCO and MSCOCOReduced lies in the above. Besides that, every operation is the same, therefore there should be no need to override the constructor (or any other function for that matter). Maybe I could modify the MSCOCO constructor to accept a dictionary and do any necessary pruning (like removing categories from the dataset) in there instead.

Suggestions are welcome as always.

[ahundt]

I'd have it be a "private" (leading underscore) member function that can be called from the constructor, but otherwise that sounds good to me.