What If We Recaption Billions of Web Images with LLaMA-3?