Text this: Detecting overfitting in Neural Networks during long-horizon grokking using Random Matrix Theory