Automatic calculation of dual function gradient

[RicardoDominguez]

Using torch and theano.

[RicardoDominguez]

Automatic gradient computation for the dual function resulted in (as tested with 065113a):

• Theano: better performance, implementation added
• Torch: significantly worse performance, implementation not added.

Moving forward, check if this is also true with torch on the GPU.