Support streaming datasets that use pathlib

[albertvillanova]

This PR extends the support in streaming mode for datasets that use pathlib.Path.

Related to: #2866.
CC: @severo

[lhoestq]

Thanks, this is very nice ! It should fix a lot of dataset in streaming mode :)

I added a few suggestions, mostly about the tests

[lhoestq]

LGTM thank you !

[severo]

I've tried #2866 again, and I get the same error.

import datasets as ds
ds.load_dataset('counter', split="train", streaming=False)

[albertvillanova]

@severo Issue #2866 is not fully fixed yet: multiple patches need to be implemented for pathlib, as that dataset uses quite a lot of pathlib functions... 😅

[severo]

No worry and no stress, I just wanted to check for that case :) I'm very happy that you're working on issues I'm interested in!