PyTorch 1.11.0 compatibility updates

[glenn-jocher]

Resolves AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor' first raised in ultralytics/yolov5#5499 and observed in all CI runs on just-released PyTorch 1.11.0.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced model loading compatibility and code simplification in YOLOv3.

📊 Key Changes

• 🧩 Unified model loading process for both fused and un-fused models.
• 🔍 Simplified evaluation mode setting for loaded models.
• ✨ Improved code readability by using conditional expressions.
• 🛠 Added compatibility updates for various PyTorch versions, including 1.6.0, 1.7.0, and 1.11.0.

🎯 Purpose & Impact

• 🚀 The purpose is to streamline the model loading process, making it more efficient and readable.
• 💡 The impact is to ensure better compatibility with different versions of PyTorch, reducing potential issues for users when loading models.
• 🔗 Users should experience a more resilient model loading, possibly resulting in fewer bugs or errors when working with different PyTorch versions.