What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness