SMART: Automatically Scaling Down Language Models with Accuracy Guarantees for Reduced Processing Fees