Create datasets package with shortcuts to acquire datasets as DataFrames

[frreiss]

Our notebooks and experiment scripts frequently repeat a pattern:

• Download a reference data set (if not already present)
• Read the data set with one of our reader functions
• Convert everything in the data set to DataFrames

We should wrap these three steps into a single function so that we and our users don't need to write this code over and over again.

Suggested API:

• Main entry point attp.dataset.download_<data set name>(), with optional arguments to specify:
  • cache directory
  • fold name
  • whether to return a DataFrame per document or a single stacked DataFrame
• Each download_<name>() function performs the following steps:
  • If the raw data set isn't present, download it
  • Convert the entire raw data set into DataFrames
  • Stack the DataFrames into a single large dataframe (add a leading column with fold name) and write this DataFrame as a single Parquet file in the cache directory
  • Use the cached Parquet file for subsequent reads of the data set
  • If the user requested a DataFrame per document, split the single large DataFrame into multiple smaller ones