refactor(config): eliminate global mutable state, wire DeerFlowContext into runtime

- Freeze all config models (AppConfig + 15 sub-configs) with frozen=True
- Purify from_file() — remove 9 load_*_from_dict() side-effect calls
- Replace mtime/reload/push/pop machinery with single ContextVar + init_app_config()
- Delete 10 sub-module globals and their getters/setters/loaders
- Migrate 50+ consumers from get_*_config() to get_app_config().xxx

- Expand DeerFlowContext: app_config + thread_id + agent_name (frozen dataclass)
- Wire into Gateway runtime (worker.py) and DeerFlowClient via context= parameter
- Remove sandbox_id from runtime.context — flows through ThreadState.sandbox only
- Middleware/tools access runtime.context directly via Runtime[DeerFlowContext] generic
- resolve_context() retained at server entry points for LangGraph Server fallback

backend/packages/harness/deerflow/config/model_config.py

@@ -12,7 +12,7 @@ class ModelConfig(BaseModel):
         description="Class path of the model provider(e.g. langchain_openai.ChatOpenAI)",
     )
     model: str = Field(..., description="Model name")
-    model_config = ConfigDict(extra="allow")
+    model_config = ConfigDict(extra="allow", frozen=True)
     use_responses_api: bool | None = Field(
         default=None,
         description="Whether to route OpenAI ChatOpenAI calls through the /v1/responses API",