Fix Qwen3.5 multimodal packing kwargs compatibility

[taking-lying-flat]

Summary

• prevent Qwen3.5 text packing kwargs (cu_seq_lens_q, cu_seq_lens_k, max_length_q, and max_length_k) from leaking into the vision tower
• register the filter as an instance-scoped, idempotent vision forward pre-hook for both dense and MoE loaders
• preserve the same packing kwargs for the language model path
• set the dense Qwen3.5 and OvisOCR2 minimum Transformers version to >=5.2.0, matching the first stable Qwen3.5 release and the existing MoE requirement

Root cause

Transformers 5.9+ forwards the model-level packing kwargs into the Qwen3.5 vision tower. The vision attention path also supplies its own sequence-length kwargs explicitly, which raises a duplicate-keyword TypeError. The hook strips only the four text packing keys at the vision boundary while leaving the original kwargs unchanged for the language model.

The previous dense/Ovis requirement (transformers>=5.0.0.dev) also admitted stable 5.0/5.1 releases that do not provide Qwen3.5. The corrected lower bound is 5.2.0.

Compatibility

Tested every published stable Transformers release in the project's supported range:

5.2.0, 5.3.0, 5.4.0, 5.5.0, 5.6.0, 5.7.0, 5.8.0, 5.9.0, 5.10.0, 5.11.0, 5.12.0, 5.12.1, 5.13.0, 5.13.1, 5.14.0, and 5.14.1.

The collision begins in 5.9.0. On 5.2-5.8 the hook remains compatible and harmless.

Local validation

• external compatibility harness (not included in this PR): all checks passed on each of the 16 supported stable Transformers releases
• Transformers 5.12.1 and 5.14.1 + Torch 2.13.0+cu130 + FlashAttention 2.8.3 + CUDA: image/video forward and backward passed
• project pre-commit checks: flake8, isort, yapf, whitespace, EOF, quote, merge-conflict, and line-ending checks passed
• git diff --check passed

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.