Rethinking Learning Rate Tuning in the Era of Large Language Models