Training neural networks faster with minimal tuning using pre-computed lists of hyperparameters for NAdamW