Move cudnn.benchmarks(True) to LoadStreams

[glenn-jocher]

Also remove unused LoadWebcam class.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvements in webcam and stream data loading for YOLOv5 inference.

📊 Key Changes

• Removed usage of torch.backends.cudnn from predict.py and detect.py, moved it to dataloaders.py.
• Deleted the LoadWebcam class, as webcam functionality is now unified with other streams in the LoadStreams class.
• Simplified stream source validation by utilizing Path reading instead of manual file checks.
• Optimized the stream update method by using cap.grab() solely rather than a read-grab combination.

🎯 Purpose & Impact

• 🚀 Performance: Centralizing cudnn.benchmark = True for consistent inference speed-up across different scripts.
• 📹 Simplicity: By removing the LoadWebcam class and consolidating webcam handling into LoadStreams, code maintenance and future updates are simplified.
• 🧼 Cleaner Code: Streamlining the method to obtain stream sources and updating frames provides cleaner and more efficient code, potentially leading to fewer bugs and easier troubleshooting.
• 🛠️ User Experience: These changes should result in a smoother and potentially faster experience for users performing detections via webcam or streamed inputs.