Prevent serialization/pickling errors when using cloud-based datasets (e.g., S3).

[SajidAlamQB]

Description

Related to: #2162

Datasets backed by cloud storage (e.g., S3 via s3fs) often hold non-picklable internal references (e.g., S3FileSystem._glob), which cause ParallelRunner to fail during multiprocessing pickling.

Potential Fix:

• Add a validation layer in ParallelRunner._validate_catalog() that warns users about cloud datasets and suggests using ThreadRunner. Detect datasets with _protocol in ("s3", "gcs", "adl", etc.) and raise a warning with suggested alternatives (e.g., ThreadRunner).

• Update documentation with guidance and caveats for cloud-based storage and multiprocessing, explain why cloud datasets are will cause issues with ParallelRunner.

[datajoely]

I wonder if we can recommend a better pickling library like cloudpickle

[astrojuanlu]

Rather than looking at this from the Runners perspective, can we maybe do so from the Datasets perspective? #1936

According to the original issue #2162 this has to do with fsspec. Custom datasets can of course introduce non-pickleable attributes, but our official ones are unpickleable because of fsspec, so it's something we can address.