fix(llm): add trust_remote_code=True to HF Transformers backend

MiniCPM3-4B and other OpenBMB models contain custom modeling code and require
trust_remote_code=True when loaded via transformers.from_pretrained().

The pipeline() API passes kwargs via model_kwargs and tokenizer_kwargs, so
add trust_remote_code=True to both to ensure the model and tokenizer load
correctly without 'custom code must be executed' errors.

hearthnet/services/llm/backends/hf_local.py

@@ -126,7 +126,9 @@ class HfLocalBackend:
             model=self._model_name,
             device=device,
             # Disable auto device_map to keep explicit CPU/GPU control
-            model_kwargs={"low_cpu_mem_usage": True},
+            # Add trust_remote_code=True for models with custom modeling code (e.g. MiniCPM3-4B)
+            model_kwargs={"low_cpu_mem_usage": True, "trust_remote_code": True},
+            tokenizer_kwargs={"trust_remote_code": True},
         )
 
     def _build_prompt(self, messages: list[dict]) -> str: