Fix docstrings and types in documentation that appears on the website

While I was working on https://github.com/huggingface/datasets/pull/5313 I've noticed that we have a mess in how we annotate types and format args and return values in the code. And some of it is displayed in the [Reference section](https://huggingface.co/docs/datasets/package_reference/builder_classes) of the documentation on the website.

Would be nice someday, maybe before releasing datasets 3.0.0, to unify it......