Computing a human-like reaction time metric from stable recurrent vision models