Support both Theano (Lasagne; Keras) and Tensorflow (Keras) backend

[Omrigan]

The idea comes from lasagne: https://github.com/Lasagne/Lasagne/issues/611

Some functions have the same functional, but 2 backends can return different format of results. However, it is not a big issue if we carefully abstract them.

The biggest advantages of adding tensorflow backend is that it significantly speed up the building and running process on CPU. Hence, it helps a lot with the development process.

I am aware that some Agentnet modules are specialised for only Lasagne, however, I think it is possible to adapt all common layers for both frameworks.

[justheuristic]

we're planning on this (and in fact could use someone help with contributions), so far you can obtain similar speed-ups by compiling with mode "FAST_COMPILE" or "DEBUG_MODE"