Reinforcement learning for automatic quadrilateral mesh generation: a soft actor-critic approach