fix(ml): load models in separate threads (#4034)

* load models in thread

* set clip mode logs to debug level

* updated tests

* made fixtures slightly less ugly

* moved responses to json file

* formatting

machine-learning/app/models/cache.py

@@ -17,7 +17,7 @@ class ModelCache:
         revalidate: bool = False,
         timeout: int | None = None,
         profiling: bool = False,
-    ):
+    ) -> None:
         """
         Args:
             ttl: Unloads model after this duration. Disabled if None. Defaults to None.