v0.2.0 - Composite DeepSeek Vision

English: Replaces the standalone vision model with one selectable DeepSeek V4 Flash + Vision composite model. Text goes directly to DeepSeek; image messages are analyzed by a hidden vision sidecar first. Existing vision-openai/GLM routes are reused automatically during migration. 中文：将原先独立的视觉模型改为一个可直接选择的 DeepSeek V4 Flash + Vision 组合模型。纯文字直达 DeepSeek，图片先由内部视觉模型识别；升级时会自动复用已有的 vision-openai/GLM 配置。