Support custom Layer types known by backend

[alexjc]

If a user specifies a Layer in the constructor that's known to the backend (lasagne) simply pass it through directly. This makes it possible to do more advanced architectures without "hitting the ceiling", at the cost of intuitiveness and simplicity.

[nikcheerla]

This would really be useful! The only reason I don't use scikit-neuralnetwork for all deep learning tasks right now is the fact that I can't easily add things like max-pooling layers or dropout layers.

EDIT: never mind, I just realized that max-pooling and dropout options exist! This would still be great in order to implement more exotic layers though.

[alexjc]

@nikcheerla The code is there, you can do things like:

layers=[Native(lasagne.layers.recurrent.LSTMLayer, <parameters>)]

Section in the documentation will follow soon.