Package reorganization

It'd be good to clarify the boundaries of dask-glm and dask-ml. My motivation is building up a set of utilities in dask-ml for working generically with dask or NumPy arrays, and dask or pandas dataframes.

I'd like to move the estimator interface over to `dask-ml` and leave all the optimization and regularization logic here. I think that `families` would stay in dask_glm.

With that reorganization, the interface is:

- dask_glm for lower-level optimization routines on large dask arrays
- dask_ml for higher-level estimators, which can work on dask, numpy, or pandas data structures.