V-GRPO: Online Reinforcement Learning for Denoising Generative Models Is Easier than You Think