Mechanistic Interpretability of LoRA-Adapted Language Models for Nuclear Reactor Safety Applications