Write method to learn softmax parameters

[durovo]

The OneVsAll softmax probability calibration method in shogun currently uses parameters learnt on a sigmoid. A method is needed that can be used to learn parameters for a softmax function instead.

The current fit_sigmoid call

shogun/src/shogun/machine/MulticlassMachine.cpp

Line 132 in d1763b8

CStatistics::SigmoidParamters params = CStatistics::fit_sigmoid(outputs[i]->get_values());

should be replaced by some method using which we can learn params for softmax.
See description of equation 7 in paper https://hal.inria.fr/inria-00103955/document.