[core] Ref counting for actor handles (#7434)

* tmp

* Move Exit handler into CoreWorker, exit once owner's ref count goes to 0

* fix build

* Remove __ray_terminate__ and add test case for distributed ref counting

* lint

* Remove unused

* Fixes for detached actor, duplicate actor handles

* Remove unused

* Remove creation return ID

* Remove ObjectIDs from python, set references in CoreWorker

* Fix crash

* Fix memory crash

* Fix tests

* fix

* fixes

* fix tests

* fix java build

* fix build

* fix

* check status

* check status

python/ray/tests/test_actor.py

@@ -106,6 +106,7 @@ def test_actor_method_metadata_cache(ray_start_regular):
 
     # The cache of ActorClassMethodMetadata.
     cache = ray.actor.ActorClassMethodMetadata._cache
+    cache.clear()
 
     # Check cache hit during ActorHandle deserialization.
     A1 = ray.remote(Actor)
@@ -532,6 +533,34 @@ def test_actor_method_deletion(ray_start_regular):
     assert ray.get(Actor.remote().method.remote()) == 1
 
 
+def test_distributed_actor_handle_deletion(ray_start_regular):
+    @ray.remote
+    class Actor:
+        def method(self):
+            return 1
+
+        def getpid(self):
+            return os.getpid()
+
+    @ray.remote
+    def f(actor, signal):
+        ray.get(signal.wait.remote())
+        return ray.get(actor.method.remote())
+
+    signal = ray.test_utils.SignalActor.remote()
+    a = Actor.remote()
+    pid = ray.get(a.getpid.remote())
+    # Pass the handle to another task that cannot run yet.
+    x_id = f.remote(a, signal)
+    # Delete the original handle. The actor should not get killed yet.
+    del a
+
+    # Once the task finishes, the actor process should get killed.
+    ray.get(signal.send.remote())
+    assert ray.get(x_id) == 1
+    ray.test_utils.wait_for_pid_to_exit(pid)
+
+
 def test_multiple_actors(ray_start_regular):
     @ray.remote
     class Counter: