Support arbitrary reference models #104
Description
Currently, the reference model is a model defined by Bambi, and the submodels are nested models of it. We could make the API more general and support arbitrary reference models. From the reference, we need an ELPD and a set of samples from the posterior predictive. In principle, the reference could be anything, but some "natural" examples are non-parametric models like PyMC-BART models and GPs (likely computed with Bambi or PyMC), but it could be something else. Notice that we still need to define a Bambi model, but this model will not be the reference one; it will be the largest possible model.