[core][cgraph] Fix doc test cgraph_overlap (#51139)

Explicitly tear down the compiled graph and kill the actors rather than relying on GC.

Also previously in Compiled Graph, actor is killed only when the actor task does not finish within timeout. This PR fixes it by always killing the actor when kill_actors=True.

Signed-off-by: Rui Qiao <ruisearch42@gmail.com>

doc/source/ray-core/doc_code/cgraph_overlap.py

@@ -68,6 +68,7 @@ def test(overlap_gpu_communication):
         assert result == [(i, shape, dtype)] * num_senders
     duration = time.monotonic() - start
     print(f"{overlap_gpu_communication=}, {duration=}")
+    compiled_dag.teardown(kill_actors=True)
 
 
 for overlap_gpu_communication in [False, True]:

python/ray/dag/compiled_dag_node.py

@@ -2144,6 +2144,13 @@ def wait_teardown(self, kill_actors: bool = False):
                     except Exception:
                         pass
 
+                if kill_actors:
+                    # In the previous loop, we allow the actor tasks to exit first.
+                    # Now, we force kill the actors if not yet.
+                    for actor in outer.worker_task_refs:
+                        logger.info(f"Killing actor: {actor}")
+                        ray.kill(actor)
+
             def teardown(self, kill_actors: bool = False):
                 with self._in_teardown_lock:
                     if self._teardown_done: