[AMD] Default Qwen3.5 architectures to text-only path

[lishuoshuo-amd]

Motivation

Qwen3.5 text architectures include a vision_config in their Hugging Face config. With the current automatic multimodal detection, Qwen3_5ForConditionalGeneration and Qwen3_5MoeForConditionalGeneration default to the multimodal serving path even when users are serving them as text-only models.
This can trigger unnecessary multimodal initialization for the text-only path. Users can still opt in explicitly with --enable-multimodal.

Modifications

• Add Qwen3_5ForConditionalGeneration and Qwen3_5MoeForConditionalGeneration to mm_disabled_models.
• Add a unit test covering the default text-only behavior and explicit multimodal opt-in.

Test plan

PYTHONPATH=python python3 -m pytest test/registered/unit/configs/test_qwen3_5_model_config.py -q