Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension