An actor-critic deep reinforcement learning approach for metro train scheduling with rolling stock circulation under stochastic demand