Use hfh for caching

## Is your feature request related to a problem?

As previously discussed in our meeting with @Wauplin and agreed on our last datasets team sync meeting, I'm investigating how `datasets` can use `hfh` for caching.

## Describe the solution you'd like

Due to the peculiarities of the `datasets` cache, I would propose adopting `hfh` caching system in stages.

First, we could easily start using `hfh` caching for:
- dataset Python scripts
- dataset READMEs
- dataset infos JSON files (now deprecated)

Second, we could also use `hfh` caching for data files downloaded from the Hub.

Further investigation is needed for:
- files downloaded from non-Hub hosts
- extracted files from downloaded archive/compressed files
- generated Arrow files

## Additional context

Docs about the `hfh` caching system:
- [Manage huggingface_hub cache-system](https://huggingface.co/docs/huggingface_hub/main/en/how-to-cache)
- [Cache-system reference](https://huggingface.co/docs/huggingface_hub/main/en/package_reference/cache)

The `transformers` library has already adopted `hfh` for caching. See:
- huggingface/transformers#18438
- huggingface/transformers#18857
- huggingface/transformers#18966