Refactor FeatureExtractror to have fit() and transform() methods

[zurk]

Ancestor issue: #387

The problem is that feature extraction is different when you train the model and when you apply it.
So it is clear that FeatureExtractor is a transformer and fit() and transform() methods should be implemented. It helps to reduce code complexity a lot. Now we have both methods implemented inside extract_features() and index_labels flag turns it to fit or transform method.

[m09]

Fit and transform is actually not the API we want. We want fit_transform & save for the training and load & transform for the testing.

[zurk]

yeah, you are right, I just decompose fit_transform to fit + transform :)