Distill Knowledge in Multi-task Reinforcement Learning with Optimal-Transport Regularization