Splitting Steepest Descent for Growing Neural Architectures