Fix mRoPE position ID crash on Qwen2-VL prompt truncation

[Mr-Neutr0n]

Summary

Fixes #441

When training Qwen2.5-VL with agent-lightning + verl, the model crashes in get_rope_index with a shape mismatch:

position_ids[..., attention_mask == 1] = llm_positions

fails because llm_positions length differs from the attention mask true-count.

Root cause: In get_train_data_batch, prompt truncation (prompt_ids[:max_prompt_length]) changes the token count, potentially removing image placeholder tokens. However, image_grid_thw is computed from the original (untruncated) image_urls list. When get_rope_index processes the truncated sequence, it finds fewer <|vision_start|><|image_pad|> regions than image_grid_thw entries, causing the position ID length to diverge from the attention mask count.

Fix: After prompt truncation, count the remaining image regions in the truncated token sequence using the same vision_start_token_id + image_token_id pattern that get_rope_index uses, and slice image_urls to match before computing image_grid_thw.

• Added _count_images_in_tokens() helper method to detect image regions in token sequences
• Modified the transition-level mRoPE code path to reconcile image_urls with truncated prompts

Test plan

• Verify Qwen2.5-VL training with prompts that exceed max_prompt_length and contain images no longer crashes in get_rope_index
• Verify Qwen2.5-VL training with prompts shorter than max_prompt_length is unaffected (no truncation, all images retained)
• Verify non-VL model training paths are unaffected (_use_mrope is False)

[Mr-Neutr0n]

Friendly bump! Let me know if there's anything I should update or improve to help move this forward.