Datasets via entry points

[seanmacavaney]

This work-in-progress allows datasets to be defined in external packages through the entry point pyterrier.dataset_provider.

The existing datasets have been moved into two providers: builtin (for default set of datasets), and irds (for the ones provided by ir-datasets).

This sets the stage for a few things:

• Allowing other packages to provide datasets without the current workaround (first import the package, then have the import add the datasets to pyterrier.datasets.DATASET_MAP).
• Dropping the dependency to ir-datasets (by allowing it to provide the pyterrier-compatible interface itself)

A side benefit is that dataset objects don't need to be created until needed. This could reduce the import time of the core pyterrier package.

The diff looks messy due to moving stuff around. The key new bits are in pyterrier/datasets/_core.py.