Faster parquet streaming + filters with predicate pushdown

[lhoestq]

ParquetFragment.to_batches uses a buffered stream to read parquet data, which makes streaming faster (x2 on my laptop).

I also added the filters config parameter to support filtering with predicate pushdown, e.g.

from datasets import load_dataset

filters = [('problem_source', '==', 'math')]
ds = load_dataset("nvidia/OpenMathInstruct-2", streaming=True, filters=filters)
first_example = next(iter(ds["train"]))
print(first_example["problem_source"])
# 'math'

cc @allisonwang-db this is a nice plus for usage in spark

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.