[ENH] online probabilistic regression meta-algorithms wishlist

[fkiraly]

Implementation wishlist for simple online probabilistic regression meta-algorithms.

• fit separate copies of regressor on the batches, and then return Mixture with weights by the number of samples in each batch. Batches under a certain size (hyperparameter) are ignored; or, collected, until a minimum size is reached (hyperparameter).
• remember a smaller bootstrap sample of size n_remember from the fit data, and pool this with data in a new batch. At each update, bootstrap so the remembered sample stays n_remember size. Predict Mixture from fit_predict on the remembered plus new sample, in update, with appropriate weights.

Further, general Bayesian algorithms will support a Bayesian update, this should be included in the API design for Bayesian estimators.