From Code to Prediction: Fine-Tuning LLMs for Neural Network Performance Classification in NNGPT