In Defense of Softmax Parametrization for Calibrated and Consistent Learning to Defer