Mix-up Consistent Cross Representations for Data-Efficient Reinforcement Learning