A Hybrid Deep-Learning Model for El Ni\~no Southern Oscillation in the Low-Data Regime