Hybrid CNN–Transformer Neural Architecture with Generative Data Augmentation and Focal Loss Optimization for Improving Accuracy in Facial Micro-Expression Recognition