Models for adaptive arithmetic coding

This is also essentially what we do in Daala, as proposed here back in 2012: https://tools.ietf.org/html/draft-terriberry-codingtools-02#...

The only difference is we don't use rANS (which didn't exist when we started), we use a different arithmetic coding technique that does not require divisions, that also handles probabilities that don't sum to a power of two, so even the traditional frequency count style of models work fine.

Glad to see others are recognizing that this is a good idea.