Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance