perform launch_if_ready asynchronously, at start of a new stack

... rather than recursively deep in a dependency processing chain

probably some change in performance?

launch_if_ready was already intended to be launched multiple times
from multiple threads. this PR might make the invocation to
launch_if_ready happen a bit later, but correctness-wise that should
be fine: a task can only become *more* ready to run, not less *ready*.

parsl/dataflow/dflow.py

@@ -614,6 +614,13 @@ def check_staging_inhibited(kwargs: Dict[str, Any]) -> bool:
         return kwargs.get('_parsl_staging_inhibit', False)
 
     def launch_if_ready(self, task_record: TaskRecord) -> None:
+        """schedules a task record for re-inspection to see if it is ready
+        for launch. The call will return immediately, asynchronous to
+        whether than check and launch has happened or not.
+        """
+        self.dependency_launch_pool.submit(self._launch_if_ready_async, task_record)
+
+    def _launch_if_ready_async(self, task_record: TaskRecord) -> None:
         """
         launch_if_ready will launch the specified task, if it is ready
         to run (for example, without dependencies, and in pending state).