Optimized the code with multi-GPU support and efficient data loading

[franciphy]

The optimization includes using tf.distribute.MirroredStrategy for multi-GPU training, which automatically handles data distribution and gradient synchronization, significantly improving efficiency across multiple GPUs. Additionally, tf.data.Dataset is used as the data loader, replacing ImageDataGenerator to enhance data loading efficiency, especially for large datasets. A learning rate scheduler is introduced through the LearningRateScheduler callback to dynamically adjust the learning rate, accelerating convergence and improving model performance. Finally, model compilation and training are performed within strategy.scope() to ensure proper synchronization and gradient updates in the multi-GPU environment.