[FT] Relax upper bound on `datasets` dependency

[lewtun]

Issue encountered

datasets version 4.0.0 was released 1 month ago and provides several nice features like multi-processing uploads for large datasets. It would be good to relax the upper bound datasets<=4.0.0 in lighteval so the latest features can be used.

Note that datasets>=4.0.0 deprecates the use of dataset loading scripts, so some care is needed to ensure that existing benchmarks like livecodebench still work. For those benchmarks, the simplest thing would be to convert the dataset loading scripts into parquet format and host under the lighteval org. This could be done with e.g.

datasets-cli convert_to_parquet <dataset_id> --trust_remote_code

cc @lhoestq for viz who has also worked a lot to ensure popular eval datasets are already migrated to Parquet

[NathanHB]

This is quite important yes, we had issues with this while uploading RULER using datasets 4.0.
This is quite prio so will do before next release.

[clefourrier]

Will be fixed by #924