Don't Do RAG: When Cache-Augmented Generation is All You Need for Knowledge Tasks