Averaged Adam accelerates stochastic optimization in the training of deep neural network approximations for partial differential equation and optimal control problems