Implement gamma-gamma model and base CLV API

[ricardoV94]

Added a notebook comparing with lifetimes: https://github.com/pymc-labs/pymmmc/blob/gamma_gamma/notebooks/clv/gamma_gamma.ipynb

And another one showing how models can be implemented by hand: https://github.com/pymc-labs/pymmmc/blob/gamma_gamma/notebooks/clv/gamma_gamma_dev.ipynb

[juanitorduz]

@ricardoV94 this is very cool! It works very nicely! 🚀 I really like the fact you explain the PYMC implementation and its limitations. If I understand correctly when conditioning on the mean purchase per user there is no way around using pm.Potential right?

For applications, you usually have a lot of users (millions). I was wondering if we should allow variational inference estimation of the parameters (maybe pathfinder from pymc-experimental?) Do you see any problem with this?

[juanitorduz]

I left some suggestions regarding documentation (doctstrings and references of the equations). Still, if you are happy we can merge this first iteration and create tickets about the docs.

[ricardoV94]

Not without a single test! :D

[larryshamalama]

Great work @ricardoV94! This API looks very nice and makes more sense than having the model as a distribution class. I left some comments and questions, as usual :)

[ricardoV94]

Ready for a final review