You Only Cache Once: Decoder-Decoder Architectures for Language Models