chore: review follow-ups for list.contains

[FBruzzesi]

Follow-ups from reviewing narwhals-dev#4001, opened against your branch so you can keep, change or drop each part.

Kernel (_arrow/utils.py)

• Consecutive small chunks are now combined into runs of at least 2^17 values before counting. Each block costs ~35 us of pyarrow calls, so inputs with many small chunks spent most of their time on that overhead. Large chunks are still sliced without copying, as before.
• Split into _list_blocks (chunking) and _list_block_contains (the count). Empty chunks are skipped, and or 0 is gone, since pc.max of a non-empty block is never null.

Your benchmark script, with two added variants and two many-chunk cases. All variants gave identical results. The script as posted had a few truncated lines, which I filled in.

Case	PR head 608d8a6	this PR
3 elements, rare (1e-5)	123 ms, 10 MB	123 ms, 10 MB
3 elements, common (30%)	130 ms, 10 MB	130 ms, 10 MB
1,000 elements, rare	108 ms, 9 MB	108 ms, 9 MB
1,000 elements, common	117 ms, 9 MB	115 ms, 9 MB
3 elements, common, chunks of 1,000 lists	627 ms, 8 MB	171 ms, 8 MB
1,000 elements, common, chunks of 10 lists	262 ms, 2 MB	138 ms, 3 MB

Dask

• With List casts supported, nw.lit(None, nw.List(...)) stopped raising and returned a String column instead, because Dask scalars drop the dtype (dask scalars don't preserve dtype dask/dask#11637, the same cause as the xfailed test_date_lit). It raises NotImplementedError again.
• The "Only pyarrow-backed lists" check is reachable (object-dtype lists with convert-string off), so it is now tested instead of marked # pragma: no cover.

Tests

• test_contains_chunked replaces the pyarrow-only chunk test. It runs on every eager backend, with tiny block sizes so that both slicing and combining run.
• test_contains_256_matches covers the uint8 to uint16 boundary, where 256 matches would wrap to 0.
• Polars 1.44 gets list.contains wrong on chunks sliced from the same array. fix: Fix list.contains on multi-chunk sliced lists pola-rs/polars#29635 fixes it but is unreleased, so the test xfails Polars>=1.44 for now.
• Simplified the Dask test constructor and the nested-dtype skips in assert_frame_equal_test.py.

AI assistance: reviewed and drafted with Claude Code. I have read the whole diff.