Practical estimation of the optimal classification error with soft labels and calibration