Add partitioning func capabilities to allow doc-types-based embedding ranking

[mskarlin]

When you have different flavors of docs (say texts of JSON/Code along with some PDF-parsed paper text), the embedding ranking won't be based on semantics alone. All of the texts that are JSON may cluster together, and while they could be highly relevant, they are sometimes embedded further away from a natural text query based on structure alone. Our summary step deals with this downstream of the embedding ranking, but some good texts will never have the opportunity to enter into the summary step in this case.

To deal with this I added an optional partitioning_fn input into the Docs object. This allows a user to make a function which will turn each Text object's Doc into an integer category. The similarity search in the retrieval step will then be split by each integer category, ranked independently, then interleaved into a final output. This way each "partition" gets a shot at being included in the summary step.

[jamesbraza]