Refactor fetching

[LilianBoulard]

The goal of this PR is to refactor the fetching system to:

• Rely on the sklearn cache implementation of fetch_openml, which makes a lot of our code outdated
• Simplify the private API
• Separate the public from the private API, since the file is about 1K lines long, making it hard to work with
• Simplify the Dataset API, so that we don't return different objects depending on the value of load_dataframe
• Align with the API of sklearn's fetchers

[GaelVaroquaux]

Hold your horses :)

We have too many things happening right now to undertake a refactor. Let's wait a few days to merge a few things and have less ongoing modifications in the code base.

[GaelVaroquaux]

What about _fetching_backend.py and _fetching_functions.py?

+1

And to further clarify, add docstrings to the file, like we do for others?

+1

[LilianBoulard]

From my quick analysis of the test error, it seems there is a difference somewhere in fetch_world_bank_indicator, which causes the indices of the two tables to mismatch. However, doing self.X.reset_index(drop=True).equals(other.X.reset_index(drop=True)) in Dataset.__eq__ doesn't work either, and I don't know why yet.

Edit: it was a dtype mismatch, thanks @Vincent-Maladiere