Matrices as arguments

[David-Estevez]

All matrices that are passed to other functions are copies of all the data of the layer/network. If you are working with relatively big networks, that may result inefficient from the memory management point of view.

Matrices with pointers/references should be implemented.

[David-Estevez]

Now matrices are loaded in input module and stored there, therefore there is only a copy of the matrix in the whole program. Neural network works with pointers to weights instead of storing the weights by itself.

Closing issue #2.