[LLM:Feature] Add Qwen3.5 video input

[Fatemanx]

Summary

Add Qwen3.5/Qwen-style video prompt support in the LLM Omni path.

This implements video decoding through OpenCV videoio, HF-style frame sampling, timestamped video prompts, video pad token replacement, temporal M-RoPE positions, and block-diagonal visual attention masks for Qwen video semantics.

Key Changes

• Add optional MNN_LLM_VIDEOIO_OPENCV support for video prompt decoding.
• Support <video>...</video> prompts with sampled video frames and timestamp groups.
• Use video_pad embeddings separately from image pad embeddings.
• Align Qwen3.5 video preprocessing with HF-style sampling and temporal visual attention.
• Add Python and C++ regression tests for video prompt semantics, sampling, and mask behavior.

Validation

• git diff --check --cached
• clang-format --dry-run --Werror ...
• python transformers/llm/export/test_qwen35_video_semantics.py
• cmake --build build_qwen35_video_e2e --target llm_demo test_omni_video_prompt -j$(nproc)
• build_qwen35_video_e2e/test_omni_video_prompt
• E2E generated 1s/25fps number videos:
  • forward 1..25 -> increasing
  • reverse 25..1 -> decreasing