ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation — ThinkLLM