Refine response model inference logic in APIRoute

- Updated the response model inference to handle cases where the model is None or not a subclass of BaseModel or a dataclass.
- Enhanced the logic to infer the response model from the endpoint function's source code when necessary, improving schema generation accuracy.

fastapi/routing.py

@@ -545,8 +545,13 @@ class APIRoute(routing.Route):
                 response_model = None
             else:
                 response_model = return_annotation
-        
-        if response_model is not None and not lenient_issubclass(response_model, BaseModel):
+                if (
+                    response_model is None
+                    or (
+                        not lenient_issubclass(response_model, BaseModel)
+                        and not dataclasses.is_dataclass(response_model)
+                    )
+                ):
                     inferred = infer_response_model_from_ast(endpoint)
                     if inferred:
                         response_model = inferred