Make Some Noise: Towards LLM audio reasoning and generation using sound tokens