Revert "Attach transforms to model"

[glenn-jocher]

Reverts #9028

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Optimizations and Updates in Classification Training and Validation.

📊 Key Changes

• Attached class names directly to the model within the training script for better clarity and consistency.
• Updated the code to work with FP16 (half-precision) inference, promoting better performance on compatible GPUs.
• Changed the way predictions and images are processed and displayed, involving half-precision adjustments and clean-up in variable usage.

🎯 Purpose & Impact

• 🔍 Streamlining the association of class names to improve the readability of the code and ensure class names are consistent throughout the training process.
• ⚡ Enhancing performance by enabling half-precision inference. This can significantly accelerate model inference on GPUs with Tensor Cores, making it more efficient.
• 🖼 Improving image handling and prediction visualization during validation, ensuring that there's a correct display of images and predictions, aligned with the newly applied half-precision computations. This can lead to a smoother debugging and validation process for developers and clearer results for users.