Deep Reinforcement Learning-Based Adaptive IRS Control with Limited Feedback Codebooks