A Tensor Low-Rank Approximation for Value Functions in Multi-Task Reinforcement Learning