[Fix] Include max_length and render options in GLM-5.3-Flash VL tokenize hash

[ShilohYu]

This fixes stale on-disk length caches for GLM-5.3-Flash VL training when max_length or the chat render options change. The cache path predicts per-sample num_tokens under the current config and persists them in a directory keyed by tokenize_fn.hash(), but _hash_str only encoded processor budgets and pack weights. Changing max_length, system_message, add_generation_prompt, enable_thinking, or reasoning_effort left the hash unchanged, so a later run reused old-config lengths: the packer budgets by old lengths while runtime truncates by new ones, and truncation landing inside a visual span degrades into silently packed fake samples.

Encode the five parameters in _hash_str, reading max_length from the constructor local since super().init stores it only afterwards. Add a regression test that classifies every init parameter as hash-affecting or not and asserts hash() changes for each one; each variant rebuilds its baseline so the shared processor cache cannot mask the assertion. Unclassified new parameters fail the test.

Stack placement: targets feat/glm53flash-f1-vl-data (#2109), so the fix can propagate into #2111. Note: existing on-disk length caches invalidate once on the next run, which is intended.

This PR also carries a small unrelated commit that repairs the lint gate on the base branch: 3de92e7 added a selector argument to tilelang_dsa_topk_indices but left tilelang_indexer_topk_from_ranges referencing it undeclared, failing ruff F821 and mypy for every stacked PR. The parameter is threaded through, with a dispatch test for both selector values.

Validation:

• pytest tests/datasets/test_glm53_vl_tokenize_fn.py: 12 passed (11 existing + 1 new), transformers 5.17.0, real GLM-5.3-Flash-25B checkpoint.
• Mutation check with the source fix stashed: the new test goes red on the first render-option variant (system_message); the pre-existing processor/pack invalidation test stays green.
• Before/after probe: the five previously unhashed configs all equaled baseline before, pairwise-distinct after; the max_pixels control changed in both.
• tests/ops/test_tilelang_dsa_topk_chunk.py::test_range_selector_dispatches: 2 passed.
• Local pre-commit on all changed files: all hooks pass, including the previously failing ruff F821 and mypy.
• No end-to-end training run or cache-directory reuse test is claimed; hash-level invalidation only.