Insights from Autoregressive vs Diffusion Large Language Models for Brain Encoding During Speech Perception