ERA: Transforming VLMs into Embodied Agents via Embodied Prior Learning and Online Reinforcement Learning