Added extra support for multilabel datasets.

[crscardellino]

Also added the possibility to load a libsvm file with column indices starting in zero (1 is still the default), like in scikit-learn's load_svmlight_file.