Formulaic and statsmodels dependencies

[jbogaardt]

chainladder currently relies on patsy for Wilkinson formulas (R-style formulas). This gets used in BarnettZehnwirth and TweedieGLM. However, patsy is no longer maintained and is instead replaced by formulaic. This has expanded features including pickling of estimators that use these formulas.

TweedieGLM could be greatly simplified (and relied on in the BootstrapODPSample if we had a few additional statistics associated with the model fit. sklearn.linear_model.TweedieRegressor only produces coefficients. statsmodels has a full GLM suite that would elminate the need for us to calculate our own statistics such as the pearson residuals or hat matrix.

All statsmodels dependencies are chainladder dependencies so this does not cause an alarming increase in dependency footprint.

[jbogaardt]

This enhancement might also make #291 viable.

[lorentzenchr]

Might be worth to consider https://github.com/Quantco/glum.