Distributed Training of Deep Neuronal Networks: Theoretical and Practical Limits of Parallel Scalability