[gboost] extend affine weak learner to structured features

[accosmin]

• the weights become a matrix
• extend the unit tests accordingly
• need to extend the dataset iterator to loop over structured features in batches