Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations