Scaling Laws for Fact Memorization of Large Language Models