Moshi: A speech-text foundation model for real time dialogue