Implement a custom loss function for sequence-to-sequence learning that ignores padding tokens during gradient computation.
