Speed up prediction by chunks

[danifranco]

We need to rethink how the final prediction is reconstructed using inference by chunks, as right now we are using so much disk (and time) with the mask that needs to be divided to take into account the overlap/padding as here.