Transfer-Tuning: Reusing Auto-Schedules for Efficient Tensor Program Code Generation