Add compressed support for multinomial logistic regression coefficients

For sparse coefficients in MLOR, such as when high L1 regularization, it may be more efficient to store coefficients in compressed format. We can add this option to MLOR and perhaps to do some performance tests to verify improvements.