feat: add multimodal fallback

[grubwithu]

[new] Multimodal fallback: auto-switch to vision model when images are detected: when a request contains images but the active model doesn't support vision (e.g. mimo-v2.5-pro), the proxy now automatically rewrites the upstream model to a vision-capable one (default mimo-v2.5) so images are processed instead of silently dropped. Toggle and model selection are in the admin UI → Codex Integration → "Thinking & Runtime Overrides" tab. Disabled by default — enable it when your workflow mixes vision and non-vision models.

---

[new] 多模态 Fallback：检测到图片时自动切换 vision 模型：当请求包含图片但当前模型不支持 vision（如 mimo-v2.5-pro）时，代理自动将 upstream model 重写为多模态模型（默认 mimo-v2.5），避免图片被静默丢弃。开关和模型选择在 admin UI → Codex 接入 →「思考与运行时覆盖」标签页。默认关闭——需要混合 vision / 非 vision 模型时开启。