Efficient Multi-task Reinforcement Learning with Cross-Task Policy Guidance