Reduce requirements by deferring import until when it's actually needed

[wengh]

This change allows use cases such as:

• reading a dataset on a PySpark version that doesn't have DataSourceArrowWriter
• writing a dataset to HuggingFace without datasets installed

[lhoestq]

Cool ! always nice to defer imports

I think we should keep those dependencies mandatory though, who does a package for which you have to install more packages to use its main feature ? ^^'

btw datasets and huggingface_hub are often installed together (datasets does require huggingface_hub)

[wengh]

I think we should keep those dependencies mandatory though, who does a package for which you have to install more packages to use its main feature ? ^^'

Yeah that makes sense. I changed pyproject to require both packages. Thanks for the correction!

[lhoestq]

lgtm !