[WIP] Improve wrapped sklearn class repr - simplify cross val for Dataset / MLDataset

[PeterDSteinberg]

• Creates an elm.pipeline.steps subpackage for wrapped sklearn transformers / estimators. This will help the repr / str and other __ private methods for the wrapped classes.
• Moves some of the logic of dask-searchcv PR 61 to Elm

[PeterDSteinberg]

@gbrener - In this PR, there are new modules in elm.pipeline.steps that map to sklearn subpackages, e.g an elm.pipeline.steps.linear_model module that has a LinearRegression wrapped for MLDataset. Here's an example of AffinityPropagration wrapped. A design question:

• Previously (e.g. elm PR 192) we had a function that used a base class Wrapped and programmatically created all these wrapped classes rather than the new elm.pipeline.steps modules. And with that approach elm.pipeline.steps.linear_model was a Namespace within the elm.pipeline.steps module (having an attribute LinearRegression). That approach had the problem described here on issue 224 where all the classes had the __name__ Wrapped because Wrapped was a class name used in the dynamic creation of wrapped classes. The good thing about the approach in master now with Wrapped is that it is more succinct, but also it means we would need to programmatically create doc strings for LinearModel as wrapped by elm.pipeline.steps.Wrapped. Most of the file diff in this PR is related to changes in elm.pipeline.steps. Do you think I should undo all the new modules in elm.pipeline.steps and fix the repr of Wrapped ? See its current definition in master here - basically it delgates everything to a class given ._cls which is a class and self is passed into _cls for many methods.