Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations