Revert to Qwen2.5-VL-7B-Instruct

Switch to Qwen3-VL-8B-Instruct

Fix model name: Qwen3-VL-30B-A3B-Instruct

Fix ranking bug, increase prompt weight, add detailed breakdown

Upgrade to Qwen3-VL-30B-A3B

Fix model loading, add skeleton placeholders

Add Qwen3-VL-30B-A3B, Compare Images, Dark UI

Add image evaluator with Soft-TIFA and VLM-as-Judge