Approximate Model-Based Shielding for Safe Reinforcement Learning