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