Support dask types

[jcrist]

This adds support for passing dask objects to fit.

Todo:

• [x] Tests

[mrocklin]

What is the behavior when I pass a dask.dataframe or dask.array? Do we not just coerce to a numpy array or pandas dataframe?

[jcrist]

No "chunked" work is done, so dask.array/dask.dataframe objects will be computed before being passed to fit. The main point of this PR is to:

• Allow inputs to be dask objects, merging in their graphs to the graph produced by the *SearchCV obect. This also allows the input data to already be on the workers (perhaps even scattered) removing the need to serialize it.
• Prevent/minimize any computation that needs to be done before building up the graph. We need to know how many splits there are beforehand. Some CV methods require knowledge of the input data before determining this (which may require computation). We specialize on the builtin CV types to avoid doing this if possible.
• Adds the refit step to the graph, removing the need for multiple calls to compute. This also allows the input data to be serialized only once (if needed).