SPEX: Scaling Feature Interaction Explanations for LLMs