DEER: A Delay-Resilient Framework for Reinforcement Learning with Variable Delays