When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method