Capture parameters from scikit-optimize for hyperparameter optimization

The Scikit-optimize `BayesSearchCV` class likely has internal functions that help it determine new parameters given scores on a set of older parameters. It would be useful to either introduce Dask to scikit-optimize or else reuse this logic and implement our own `BayesSearchCV`. This could be used with standard `fit` style algorithms or with `Incremental` if we have a nice early-stopping criterion.

This came out of conversation with @ogrisel

Start by inspecting scikit-optimize's BayesSearchCV entry point and the existing Dask-ML estimator and Incremental interfaces. Compare reusing its parameter-selection logic with introducing a Dask-native BayesSearchCV, and define done as a working fit-style integration with a clear early-stopping path if Incremental is supported.

Written by the indexing model from the issue text.