On the Improvement of Generalization and Stability of Forward-Only Learning via Neural Polarization