Module namespace cleanup for v2.0

[mariosasko]

This is an attempt to make the user-facing datasets' submodule namespace cleaner:

In particular, this PR does the following:

• removes the unused zip_nested and flatten_nest_dict and their accompanying tests
• removes pyarrow from the top-level namespace
• properly uses __all__ and the from <module> import * syntax to avoid importing the <module>'s submodules
• cleans up the utils namespace
• moves the temp_seed context manage from datasets/utils/file_utils.py to datasets/utils/py_utils.py

[severo]

will it solve https://github.com/huggingface/datasets-preview-backend/blob/4c542a74244045929615640ccbba5a902c344c5a/pyproject.toml#L85-L89?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[mariosasko]

@severo No, this PR doesn't fix that issue in the current state. We can fix it by adding __all__ to datasets/__init__.py and datasets/formatting/__init__.py. However, this would require updating __all__ for each new function/class definition, which could become cumbersome, and we can't do this dynamically because mypy is a static type checker.

@lhoestq @albertvillanova WDYT?

[lhoestq]

Thank you ! This is much cleaner

[lhoestq]

I think it's fine if we have tqdm directly in utils with disable_progress_bar

[mariosasko]

Is it okay if I address this in a separate PR before the release (I'll make disable_progress_bar public again + add enable_progress_bar, enable_caching and disable_caching) to close #3586?

[lhoestq]

Sure !

[lhoestq]

Here ArrowWriter and ArrowReader are removed from the top level module. I've seen that the lxmert example in transformers needs datasets.ArrowWriter, so I'm not sure if we should remove them

[mariosasko]

They are removed because I see them as our internals. I'll open a PR in Transformers to update their paths (used here and here).

[lhoestq]

Ok :)

[lhoestq]

Feel free to merge this one if it's good for you :)