Automatically Interpreting Millions of Features in Large Language Models