Certifiably Robust Reinforcement Learning through Model-Based Abstract Interpretation