EMA FP32 assert classification bug fix

[glenn-jocher]

Resolves v6.2 classification training bug around FP16 EMA updates (bad!)

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements to model compatibility, export options, and variable handling in YOLOv5.

📊 Key Changes

• Changed class names from a list to a dictionary in multiple scripts for consistency.
• Updated the export.py default export formats by removing 'onnx' from the default settings.
• Removed an unused import (Conv) in experimental.py for cleaner code.
• Added model compatibility updates in attempt_load function, including stride adjustments and conversion of names to dictionaries for better handling.
• Revised the ModelEMA class to omit unnecessary FP16 conversion.
• Ensured that all EMA updates occur in FP32 to maintain consistency within the model.

🎯 Purpose & Impact

• 📈 Improved compatibility and maintenance by standardizing class names to dictionaries, making the codebase more uniform and easier to manage.
• 🔧 Adjusted the default export formats, optimizing the focus for the most commonly used export types, potentially simplifying the export process for users.
• 🛠️ Code cleanup and compatibility adjustments aid in a smoother experience when loading models and contribute to a reduction in possible future issues due to data type discrepancies.
• 🤖 The general codebase cleanup and performance optimization may offer a more streamlined development experience while maintaining backward compatibility with existing models.
• 💾 Consistency in using FP32 for EMA updates improves the reliability and stability of the training process.