refactor: simplify multimodal preprocessing expansion

[CUHKSZzxy]

Summary

• Split bundled multimodal expansion into image, video, and audio helpers.
• Make bundled video expansion explicit for per-video and per-frame offset layouts.
• Keep Qwen3 Omni whole-video items and Qwen3VL/Qwen3.5 per-frame items aligned with the one-span MultiModalData contract.

Validation

• Focused VLM preprocessing tests passed.
• Qwen3 Omni processor tests passed.
• Qwen3.5 single-video pipeline smoke passed with a local cached model/media setup.
• Pre-commit hooks passed for the touched file.

Assistance

Assisted with Codex + GPT-5.5 xHigh Fast

[Copilot]

Pull request overview

Refactors get_expanded_mm_items multimodal preprocessing to split bundled HF-processor outputs into dedicated image/video/audio expansion helpers, and clarifies the intended offset layouts for whole-video vs per-frame video items to keep MultiModalData’s one-span-per-item contract.

Changes:

• Added _expand_bundled_image_items, _expand_bundled_video_items, and _expand_bundled_audio_items helpers and delegated bundled expansion to them.
• Made video expansion distinguish “one offset per video” vs “one offset per frame” paths more explicitly.
• Simplified non-bundled video_second_per_grid scalar extraction.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.