Forward iter_arrow, is_typed and features from RepeatExamplesIterable

[ebarkhordar]

RepeatExamplesIterable never forwarded iter_arrow, is_typed or features to the iterable it wraps, so .repeat() stranded an Arrow-capable child behind a non-Arrow parent: iteration fell back to the per-example path and the dtypes were re-inferred from Python objects.

This copies the three-property block and the _iter_arrow loop from SkipExamplesIterable and TakeExamplesIterable, which sit either side of it. _iter_arrow mirrors __iter__ exactly, including the per-pass reset of the child state dict, so state_dict()/load_state_dict() behaves the same on both paths.

Verification

On main at b7cb10b, Python 3.12, pyarrow 25.0.0, numpy 2.5.1:

• The three new tests fail before the change and pass after. test_iterable_dataset_repeat_keeps_numpy_dtypes reports int64/float64/int64 before against a declared int32/float32/uint8; test_repeat_examples_iterable_arrow asserts iter_arrow is not None and runs both assert_load_state_dict_resumes_iteration and assert_load_state_dict_resumes_arrow_iteration.
• tests/test_iterable_dataset.py goes from 441 passed to 445 passed, with the same 4 pre-existing test_interleave_dataset_with_sharding failures on both sides (a missing optional dependency in my environment, unrelated to this change).
• Arrow-format iteration over 100k rows drops from 6.644s to 0.469s, which is parity with .take(100000) at 0.465s. list(ds.repeat(2)) is unchanged.
• ruff format --check and ruff check are clean on both files.

What I did not check: the torch formatter, and non-scalar feature types such as Image. The dtype claim above is scoped to scalar Value columns under numpy format.

Fixes #8394