[unclean] First proposition of 1st-order sobolev, working. With very few

points, the training converges on both f and df and provide a much
better overall reconstruction of the original function than with f
alone.