Training and Comparative Evaluation of KAN and Shallow MLP Models on CIFAR-10 Using Augmentation, Model Checkpointing, and Test Set Analysis