Stable Autoregressive Speech Generation with Low-Frame-Rate High-Dimensional Continuous Tokens