Nflwin

This release provides Python 3 capability and fixes CI.

The initial release of NFLWin. Contains a working Win Probability model, as well as utilities to streamline querying [nfldb](https://github.com/BurntSushi/nfldb) (if the user has it installed) and building new models.

Biggest change is that P-values for validation are now replaced with the max deviation from perfect prediction and the total area of the residuals between predicted and expected. Hopefully this will be more robust for diagnosing if a new model is working better than the old one.

The original version required column names to be specified as keyword arguments and meant that any time any changes to the default model columns happened the model API would break. So that was bad. Now column names and definitions are stored in a dictionary that's a class attribute, so updating the API just means editing the keys/values of that dictionary.

Minor bugfixes and production testing of the `increment_version.sh` script.

Releasing now to ensure that stuff like PiPy works.