refactor: finetune through LLMClient interface + get_running_loop

- Add submit_finetune and get_finetune_status abstract methods to LLMClient base
- Implement both methods in ZhipuAIClient using asyncio.get_running_loop()
- Rewrite finetune_service to call llm.submit_finetune / llm.get_finetune_status
  instead of accessing llm._client directly, restoring interface encapsulation
- Replace asyncio.get_event_loop() with get_running_loop() in ZhipuAIClient._call
  and all four methods in RustFSClient (deprecated in Python 3.10+)
- Update test_finetune_service to mock the LLMClient interface methods as AsyncMocks
- Add two new tests in test_llm_client for submit_finetune and get_finetune_status

app/clients/llm/base.py

@@ -9,3 +9,11 @@ class LLMClient(ABC):
     @abstractmethod
     async def chat_vision(self, model: str, messages: list[dict]) -> str:
         """Send a multimodal (vision) chat request and return the response content string."""
+
+    @abstractmethod
+    async def submit_finetune(self, jsonl_url: str, base_model: str, hyperparams: dict) -> str:
+        """Submit a fine-tune job and return the job_id."""
+
+    @abstractmethod
+    async def get_finetune_status(self, job_id: str) -> dict:
+        """Return a dict with keys: job_id, status (raw SDK string), progress (int|None), error_message (str|None)."""