Foundational Autoraters: Taming Large Language Models for Better Automatic Evaluation