Fix save_to_disk writing no shard for a dataset with no rows

[LeSingh1]

save_to_disk computes the number of shards from the estimated size and then clamps it to the number of rows (added in #7912):

num_shards = min(len(self.data), num_shards)

For a dataset with no rows that gives 0 shards. No .arrow file is written, state.json gets an empty "_data_files", and loading it back fails:

ds = Dataset.from_dict({"col_1": []}, features=Features({"col_1": Value("int64")}))
ds.save_to_disk(path)   # succeeds, writes only state.json + dataset_info.json
load_from_disk(path)    # IndexError: list index out of range

save_to_disk reports success, so the data loss is silent until the load fails later.

It is reachable from an ordinary pipeline rather than only from a hand-built empty dataset:

ds.filter(lambda x: False).flatten_indices().save_to_disk(path)

and the same applies through DatasetDict.save_to_disk when one split filters down to nothing.

Passing num_shards=1 explicitly already worked, so only the automatic path was affected. This clamps the computed value to at least one shard.

Added a regression test asserting the shard file exists and the dataset round-trips; it fails on main with assert (dataset_path / "data-00000-of-00001.arrow").exists() and passes with the change. tests/test_arrow_dataset.py and tests/test_dataset_dict.py pass (469 passed, 26 skipped).