Enhancing In-context Learning via Linear Probe Calibration