Implement feature extraction

[accosmin]

• interface to access samples: tabular (feature selection via gboost), dense1d (linear models)
• normalize inputs and fill missing inputs at the model level in one place
• variations for 1D data (e.g. tabular datasets): unit (as-is), quadratic terms, linear and log-scaled normalization, quantization
• variations for 2D data (e.g. image recognition): gradients (orientation + magnitude), HoG, LBP variations, Haar, Gabor filters, random kitchen sinks, fast food etc.
• extend app/bench_gboost|linear to experiment with various features and datasets
• extend app/bench_gboost|linear to export training results (csv, models) and add python scripts to visualize results