Switch from cached_download to hf_hub_download in remaining occurrences

[Wauplin]

(somehow related to huggingface/huggingface_hub#2317)

This PR get rids of the remaining cached_download in favor of the more robust hf_hub_download. Cached_download is quite legacy (replaced since Aug 2022) and I'd like to completely drop it at some point. The core part of transformers's logic made the switch quite some time ago already.

Main change is basically to go from

json.load(open(cached_download(hf_hub_url(repo_id, filename, repo_type="dataset")), "r"))

to

json.loads(Path(hf_hub_download(repo_id, filename, repo_type="dataset")).read_text())

I used Path(...).read_text() to avoid opening the file without closing it.

Note: there is still one occurrence of cached_download in ./src/transformers/models/deprecated/van/convert_van_to_pytorch.py but since it is in a deprecated/ folder, I did not bother maintaining it.

[amyeroberts]

Thanks for updating all of these!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

That was quick! Thanks for the review 🤗