[Serve] fix flaky test_autoscaling_policy (ray-project#37527)

Calling `.clear()` will free up the memory and can cause issue in raylet. Refactored `shutdown_cached_handles()` method for iterating through all cached handles and remove their reference to safely garbage collect the handles (which is what's done when calling `serve.shutdown()`).

python/ray/serve/_private/client.py

@@ -94,16 +94,22 @@ def __del__(self):
     def __reduce__(self):
         raise RayServeException(("Ray Serve client cannot be serialized."))
 
+    def shutdown_cached_handles(self):
+        """Shuts down all cached handles.
+
+        Remove the reference to the cached handles so that they can be
+        garbage collected.
+        """
+        for cache_key in list(self.handle_cache):
+            del self.handle_cache[cache_key]
+
     def shutdown(self, timeout_s: float = 30.0) -> None:
         """Completely shut down the connected Serve instance.
 
         Shuts down all processes and deletes all state associated with the
         instance.
         """
-
-        # Shut down handles
-        for k in list(self.handle_cache):
-            del self.handle_cache[k]
+        self.shutdown_cached_handles()
 
         if ray.is_initialized() and not self._shutdown:
             try:

python/ray/serve/tests/conftest.py

@@ -83,7 +83,7 @@ def serve_instance(_shared_serve_instance):
     # Clear all state between tests to avoid naming collisions.
     _shared_serve_instance.delete_deployments(serve.list_deployments().keys())
     # Clear the ServeHandle cache between tests to avoid them piling up.
-    _shared_serve_instance.handle_cache.clear()
+    _shared_serve_instance.shutdown_cached_handles()
 
 
 def check_ray_stop():