Assessing the Macro and Micro Effects of Random Seeds on Fine-Tuning Large Language Models