Update add_vision_id logic

[hhaAndroid]

Previously, add_vision_id only took effect when the number of images or videos was greater than 1. Now it has been changed to always take effect.

Note: Breaking Change

[Copilot]

Pull request overview

This PR updates the add_vision_id logic to always take effect when enabled, removing the previous restriction that only applied it when multiple images or videos were present. This is a breaking change that aligns the behavior with HuggingFace's official implementation.

Key Changes:

• Vision IDs (e.g., "Picture 1:", "Video 1:") are now always added when add_vision_id=True, even for single images/videos
• Global counters (total_img_cnt, total_video_cnt) now track vision element numbering across entire conversations instead of per-message
• Comprehensive test coverage added for both single and multi-image/video scenarios with the new behavior

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
xtuner/v1/datasets/mllm_tokenize_fn/qwen3_vl_tokenize_fn.py	Added global total_video_cnt counter and updated video token replacement to always apply vision IDs when enabled; removed outdated comment explaining the old behavior
xtuner/v1/datasets/mllm_tokenize_fn/base_mllm_tokenize_fn.py	Added global total_img_cnt counter and updated image token replacement to always apply vision IDs when enabled
tests/resource/mllm_sft_multi_image_example_data.jsonl	Added new test case (id: 11) with images distributed across multiple messages to test global counter behavior
tests/datasets/test_qwen3_vl_tokenize_fn.py	Parametrized tests to cover both add_vision_id=True and False cases; updated test logic to properly validate single-image scenarios and removed conditional logic that disabled vision IDs for single videos

---

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