Dimer-Enhanced Optimization: A First-Order Approach to Escaping Saddle Points in Neural Network Training