Analyzing the Effects of Supervised Fine-Tuning on Model Knowledge from Token and Parameter Levels