Collect case studies to motivate future development

To correctly drive development in Dask-ML we need problems to optimize. These will help us to design and prioritize work to help real-world use cases.

A case study should operate on real data at both a local and distributed scale. They need to balance between representing pragmatic workloads and being simple enough to integrate into development practice.

- Ideally they would be executable on a subset of the data in a few minutes on a laptop, but can also scale out to run meaningfully for a long while on a distributed cluster.
- They should stress a few parts of the typical scikit-learn style process, including pre-processing and training
- They don't necessarily need to use dask-ml functionality to start out. Something that just uses scikit-learn today but that aspires towards parallel computing would be welcome
- They should be somewhat self contained. If data should be downloaded or lightly pre-processed in some way that's expected. This should be either in an idempotent cell of a notebook like `if not os.path.exists('mydata.csv'): download('http://.../mydata.csv')`, or in an attached script.
- They should be easy to understand, and not include too many domain specific details. They will need to be manipulated by developers who are not well versed in the domain of the problem and also not well versed in the practice of machine learning.

Note that these differ from the examples in the [dask-examples](https://github.com/dask/dask-ml/issues/302) repository, which are genearlly simpler, on smaller data, and primarily designed for education.

Case studies should probably live as notebooks within this repository to start, possibly with .py scripts or .md instructions if necessary for data preparation.

Eventually it would be useful to test case studies on small datasets frequently in CI systems and benchmark them with larger datasets infrequently, but this sort of infrastructure can come later (though it's good to have in mind)