Verification-Guided Shielding for Deep Reinforcement Learning