Dynamic Model Predictive Shielding for Provably Safe Reinforcement Learning