Score as Action: Fine-Tuning Diffusion Generative Models by Continuous-time Reinforcement Learning