Add a Local Deep Kernel Learning trainer

[yaeldekel]

http://manikvarma.org/pubs/jose13.pdf

[justinormont]

Great idea! I have seen LDSVM perform well on text (ngram based featurization).

Background:
LDSVM uses trees as its SVM kernel to create a non-linear binary learner which is fast at prediction time.

Slides:
http://manikvarma.org/talks/ICML13.pdf

[codemzs]

Please do not start working on this until this is triaged.