[New Feature] extending SoftmaxLoss.py to accept for class probabilities in target

[i-plusplus]

SoftmaxLoss loss currently accepts class labels[0-n) in the target.

One important use case is to fit a model on observed data with no one category but a mix of multiple classes. Changing this class to handle probabilities along with classes will cater to this use case.

Code change also seems simple. I can do the code change, if this feature is accepted.