Cast multi-dimensional numpy arrays when inferring IterableDataset features

[vineethsaivs]

Summary

IterableDataset._resolve_features() crashes when a map function yields multi-dimensional numpy arrays:

from datasets import Dataset
import numpy as np
ds = Dataset.from_dict({"a": [[[1, 2, 3], [1, 2, 3]]]}).to_iterable_dataset().map(lambda x: {"a": [np.array(x["a"])]})
ds._resolve_features()
# pyarrow.lib.ArrowInvalid: Can only convert 1-dimensional array values

Feature inference goes _resolve_features() -> _infer_features_from_batch(...) -> pa.Table.from_pydict(batch), and PyArrow's from_pydict only accepts 1-dimensional array values, so any 2D+ numpy array in a column crashes it.

Every other place in iterable_dataset.py that builds an Arrow table from raw examples first routes the data through cast_to_python_objects(..., only_1d_for_numpy=True) (lines 173 and 184), which recursively turns multi-dimensional numpy/tensor values into nested lists of 1-D arrays that PyArrow accepts. The non-streaming Dataset/ArrowWriter path does the same. _infer_features_from_batch was simply missing that cast.

Fix

Route the batch through cast_to_python_objects(batch, only_1d_for_numpy=True) before from_pydict. The helper is already imported in this module, and it is a no-op for ordinary Python data, so plain batches infer identically.

Fixes #7100.

Test

tests/test_iterable_dataset.py::test_iterable_dataset_resolve_features_from_multidim_numpy resolves the features of an IterableDataset whose map yields multi-dimensional numpy arrays. It raises ArrowInvalid before the fix and passes after (resolving to a nested List(...)).