[Datasets] Create Dataset download component.

[cweill]

We need a way for TFDS to lazily download and prepare datasets, so that it only runs at pipeline execution time instead of during DAG construction time.

[nikhil-dce]

Do you think we still need this? We can change the TFDS data_dir (e.g. https://github.com/googleinterns/nitroml/blob/nitroml_cloud/examples/titanic_benchmark.py#L50)