INTERS: Unlocking the Power of Large Language Models in Search with Instruction Tuning