feat(provider): add plugin system for service-specific model discovery

Model discovery currently assumes every provider exposes all models via
GET {baseURL}/models. Providers like Alibaba Cloud Bailian serve video,
3D and some image models through dedicated endpoints, so 21 models across
5 families (HappyHorse / Wan / Kling / PixVerse / Vidu / Tripo) are never
discoverable - a functional blocker for a video creation workbench.

This introduces a ModelDiscovery plugin interface with a registry that
matches vendor plugins by base URL and falls back to the standard OpenAI
implementation, keeping core discovery logic vendor-agnostic.

- add provider.ModelDiscovery interface and plugin registry
- add openai plugin (standard behaviour, default fallback)
- add bailian plugin (standard + 21 extended multimodal models)
- gate behind ENABLE_PROVIDER_PLUGINS, default false
- register plugins lazily from service layer to avoid import cycle

Verified locally: flag off -> 238 models (unchanged); flag on -> 257
models (238 + 19 net new after dedup). OpenAI and Gemini channels
unaffected in both modes.

docker-compose.local.yml

@@ -18,6 +18,7 @@ services:
       CANVAS_ALLOW_PRIVATE_UPSTREAMS: ${CANVAS_ALLOW_PRIVATE_UPSTREAMS:-false}
       CANVAS_ALLOWED_PRIVATE_UPSTREAM_HOSTS: ${CANVAS_ALLOWED_PRIVATE_UPSTREAM_HOSTS:-}
       CANVAS_CORS_ORIGINS: ${CANVAS_CORS_ORIGINS:-http://localhost:3000}
+      ENABLE_PROVIDER_PLUGINS: ${ENABLE_PROVIDER_PLUGINS:-false}
       GIN_MODE: release
     volumes:
       - backend-data:/data