What does this PR do?

Fixes IterableDataset.skip(n) for streaming datasets when the underlying iterable uses Arrow batches and n skips one or more complete Arrow blocks.

Previously, after a full Arrow block was counted as skipped, _iter_arrow() continued into the partial-slice branch and yielded rows from a block that should have been fully skipped.

What was the issue?

SkipExamplesIterable._iter_arrow() handles skipping in two cases:

1. the current Arrow table is fully skipped
2. only the beginning of the current Arrow table is skipped, and the rest is yielded

The bug was that case 1 did not stop after marking the table as skipped. So the same table then fell through into case 2.

In other words, a table could first be counted as "already skipped", but then still be sliced and yielded.

For example, if Arrow tables have 4 rows each and we call skip(6):

• table 1 has rows [0, 1, 2, 3] and should be fully skipped
• table 2 has rows [4, 5, 6, 7], so only [4, 5] should be skipped and [6, 7] should be yielded

Before this PR, after table 1 was counted as skipped, the code kept processing table 1 and yielded part of it. This is why skipped rows could appear in the output.

This PR adds continue after a table is fully skipped, so the code moves directly to the next Arrow table.

Tests

• PYTHONPATH=src pytest tests/test_iterable_dataset.py::test_skip_arrow_examples_iterable -q

The regression test covers skipping within a block, exactly one block, across blocks, and beyond the dataset length.