Towards Auto-Regressive Next-Token Prediction: In-Context Learning Emerges from Generalization