Choose svm or (an augmented) svmlin

[kbenoit]

textmodel_svm (based on LiblineaR::LiblineaR()) is much slower than textmodel_svmlin() (based on RSSL::svmlin()) but has some nifty features, such as a translation into multi-class outcomes and predicted probabilities.

If predictions are the same, we should add predicted probabilities and multi-class prediction for the svmlin model and use that.