[gboost] convolutional filters as weak learners for structured features

[accosmin]

• fit various filter sizes (e.g. from 3x3 to 15x15) for any given structured feature of a large enough size
• extend the unit test accordingly
• need to extend the dataset iterator to loop over structured features in batches