Re-Revert "[Core] zero-copy serializer for pytorch (#12344)" (#12478)

* [Core] zero-copy serializer for pytorch (#12344)

* zero-copy serializer for pytorch

* address possible bottleneck

* add tests & device support

(cherry picked from commit 0a505ca83d)

* add environmental variables

* update doc

python/ray/serialization.py

@@ -26,6 +26,7 @@ from ray._raylet import (
     MessagePackSerializedObject,
     RawSerializedObject,
 )
+from ray import serialization_addons
 
 logger = logging.getLogger(__name__)
 
@@ -155,6 +156,7 @@ class SerializationContext:
         # Because objects have default __reduce__ method, we only need to
         # treat ObjectRef specifically.
         self._register_cloudpickle_reducer(ray.ObjectRef, object_ref_reducer)
+        serialization_addons.apply(self)
 
     def _register_cloudpickle_reducer(self, cls, reducer):
         pickle.CloudPickler.dispatch[cls] = reducer