Restructure Parquet dataset download and processing

[daverigby]

Dataset from datasets.py handles multiple aspects of Parquet datasets:

• Downloading Parquet Datasets (consisting of multiple files in passages and queries set)
• Maintaining a local cache of previously downloaded files
• Reading records from Parquet files
  • For passages, this is done by iterating over batches
  • For queries, this is done by reading the entire query set into memory.

Additionally there's functionality which is currently unused:

• Generating queries by sampling passages (where a dataset doesn't have a queries set).

All of this leads to a class which is more complex than we need it to be, and can be harder to test.

Look at how we can restructure to simplify - potentially splitting the class into multiple independent parts (Downloading, caching, reading), and removing functionality we don't need.