Torch CUDA synchronize update

[glenn-jocher]

Fix for #1816. Torch CUDA synchronization is not attempted now on CUDA-enabled machines when --device cpu is requested.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved CUDA device handling and reproducibility configuration in PyTorch utilities.

📊 Key Changes

• Unified assignment of cudnn.deterministic and cudnn.benchmark values for clarity.
• Simplified CPU request handling in select_device by setting environment variable CUDA_VISIBLE_DEVICES to '-1'.
• Enhanced device selection string output to be more informative and user-friendly.
• Ensured batch_size is a multiple of GPU count, raising an error for incompatibility.
• Optimized GPU device properties fetching and logging.
• Made the time_synchronized function more concise.

🎯 Purpose & Impact

• Streamlining CUDA configurations improves code clarity and reproducibility when seeding for randomness 🔁.
• Clarifications in device selection facilitate better handling of CPU-only requests and ensure better error messaging 🖥️.
• More informative logging assists users in understanding their device utilization and system capabilities 📈.
• Batch size check against GPU count helps prevent runtime errors due to configuration mismatches 🛠️.
• Cleaner codebase with compact and improved readability, aiding maintenance and development 🧹.

These changes help provide a smoother experience for users working with different devices and configurations, aiming to improve the overall reliability and performance of machine learning workflows running on the Ultralytics YOLOv5 project.

[glenn-jocher]

PR is verified, all tests passing: single GPU, multi-GPU, specific multi-GPU order, CPU.