Fix silent dataset truncation in batched map with mismatched empty column (#6879)

[florian6973]

What does this PR do?

Fixes #6879.

Dataset.map(fn, batched=True) silently truncates the dataset to 0 rows (instead of raising) when the mapping function empties an existing column while another column stays non-empty.

Reproducer (from the issue)

import datasets
data = datasets.Dataset.from_dict({"test": [1]})

def mapping_fn(examples):
    return {"test": [], "y": [1]}

data = data.map(mapping_fn, batched=True)
print(len(data))   # 0  -> silent data loss

Returning an empty list for a new column (e.g. {"x": []}) already raises pyarrow.lib.ArrowInvalid, so the two cases were inconsistent.

Root cause

ArrowWriter._write_batch skips a batch that "appears empty", but the emptiness check only looked at the first column:

if batch_examples and len(next(iter(batch_examples.values()))) == 0:
    return

When a batched map overwrites an existing column with [], that column stays first in the merged dict, so the whole length-mismatched batch was dropped instead of erroring.

Fix

Skip the batch only when every column is empty. Genuinely empty batches (e.g. a batch that filters out all its rows) are still ignored, while mismatched batches now fall through to pa.Table.from_arrays, which raises the same informative ArrowInvalid already produced for the new-column case.

if batch_examples and all(len(col_values) == 0 for col_values in batch_examples.values()):
    return

Behavior

	issue reproducer
before	no exception, len(data) == 0
after	ArrowInvalid: Column 1 named y expected length 0 but got length 1

Normal batched map and legitimate "all columns empty" batches are unchanged.

Tests

Added test_write_batch_with_mismatched_empty_column_raises in tests/test_arrow_writer.py. The existing test_write_batch already covers the legitimate all-empty-batch skip, so that path stays protected.

ruff check / ruff format --check pass; tests/test_arrow_writer.py passes locally.