Layer-Parallel Training with GPU Concurrency of Deep Residual Neural Networks via Nonlinear Multigrid