RL-CFR: Improving Action Abstraction for Imperfect Information Extensive-Form Games with Reinforcement Learning