Revisiting subword tokenization: A case study on affixal negation in large language models