fix(qwen_image): forward multimodal token type ids

[llc1123]

Summary

• forward processor-provided mm_token_type_ids to the Qwen2.5-VL text encoder
• keep older Transformers processors compatible by adding the argument only when present
• add a weight-free regression covering both processor shapes

Transformers 5.15 uses mm_token_type_ids to compute Qwen2.5-VL multimodal 3D position IDs. The current image-edit path manually selects processor outputs and drops this field, which silently falls back to sequential text-only positions.

Fixes #1416

Validation

• python -m unittest discover -s tests -p 'test_qwen25_vl_input_forwarding.py' -v
• pre-commit run --files lightx2v/models/input_encoders/hf/qwen25/qwen25_vlforconditionalgeneration.py tests/test_qwen25_vl_input_forwarding.py