Save Dataset as Sharded Parquet

### Feature request

`to_parquet` currently saves the dataset as one massive, monolithic parquet file, rather than as several small parquet files. It should shard large datasets automatically.

### Motivation

This default behavior makes me very sad because a program I ran for 6 hours saved its results using `to_parquet`, putting the entire billion+ row dataset into a 171 GB *single shard parquet file* which pyarrow, apache spark, etc. all cannot work with without completely exhausting the memory of my system. I was previously able to work with larger-than-memory parquet files, but not this one. I *assume* the reason why this is happening is because it is a single shard. Making sharding the default behavior puts datasets in parity with other frameworks, such as spark, which automatically shard when a large dataset is saved as parquet.

### Your contribution

I could change the logic here https://github.com/huggingface/datasets/blob/bf6f41e94d9b2f1c620cf937a2e85e5754a8b960/src/datasets/io/parquet.py#L109-L158
to use `pyarrow.dataset.write_dataset`, which seems to support sharding, or periodically open new files. We would only shard if the user passed in a path rather than file handle.