Towards A Unified View of Answer Calibration for Multi-Step Reasoning