SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection