When Does Bigger Help? A Controlled Study of LLM Scale for Ontology Learning