This PR adds the mixture model, perturbing a parametric model using a nonparametric one.

This implementation seems to be quite stable numerically (with the Dirichlet-multinomial reducing to the multinomial for large enough total concentration parameter. At least in some test settings, with some departures visible for more complex problems involving many data points), although the parametric model requires further tests before it can be fully trusted.

The next steps would be to create a Snakemake workflow with evaluation of all of these methods, allowing one to evaluate the predictive performance (rather than analyzing coefficients).