Deep Reinforcement Learning for Robotic Manipulation under Distribution Shift with Bounded Extremum Seeking