MetaCURL: Non-stationary Concave Utility Reinforcement Learning