Fast dataset iter

[mariosasko]

Use pa.Table.to_reader to make iteration over examples/batches faster in Dataset.{__iter__, map}

TODO:

• benchmarking (the only benchmark for now - iterating over (single) examples of bookcorpus (75 mil examples) in Colab is approx. 2.3x faster)
• check if iterating over bigger chunks + slicing to fetch individual examples in _iter yields better performance

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[mariosasko]

I ran some benchmarks (focused on the data fetching part of __iter__) and it seems like the combination table.to_reader(batch_size) + RecordBatch.slice performs the best (script with the results). I think we can choose (implicit) batch_size=10 in the final implementation to avoid having problems with fetching large examples.

[lhoestq]

Nice thank you ! The benchmarks are super helpful as well, I think you can link to them in comments in the code