Optimizing In-Context Learning for Efficient Full Conformal Prediction