Lazy fitting

Should we add a `compute` keyword to all the `fit` / `partial_fit` on estimators we implement?

This would aid with

- debugging the graphs we build
- Better scheduling (any examples here?)

This is somewhat difficult. As @jcrist summarized in https://github.com/dask/dask-searchcv/issues/19

> Lazy evaluation conflicts with scikit-learn's api, and is tricky to support here in a way that is both intuitive and robust. For some configurations there is also the need to call get multiple times, which prevents lazy evaluation.

xref https://github.com/dask/dask-ml/pull/259