A Study on the Calibration of In-context Learning