Can we integrate spatial verification methods into neural-network loss functions for atmospheric science?