Align Visualized-BGE multimodal attention masks

[zjn20030811]

Summary

• align Visualized-BGE multimodal attention masks with the [CLS, image tokens, text] embedding layout
• preserve the prompt mask dtype/device and validate mask/length inputs
• add regression tests for padded prompts, boolean masks, and empty image-token sequences

Problem

encode_mm prepends image tokens after the prompt CLS embedding and removes the original prompt CLS from the remaining text sequence. Its mask was built as [image mask, full prompt mask], shifting every mask entry by one position. As a result, padding could be attended to while valid text or image positions were masked, and mean pooling used the wrong validity positions.

The new helper builds [prompt CLS mask, image mask, prompt remainder mask], exactly matching the sequence passed to the encoder. Existing model weights and APIs are unchanged.

Validation

• python -m pytest -q tests/test_visual_bge_attention_mask.py (4 passed)
• python -m compileall -q research/visual_bge/visual_bge/modeling.py