PARL: A Dialog System Framework with Prompts as Actions for Reinforcement Learning