zero-mAP final fix: Revert "torch.empty() for speed improvements"

[glenn-jocher]

Reverts #9025

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Transition from uninitialized to zero-initialized tensors in various components.

📊 Key Changes

• Replaced torch.empty with torch.zeros in multiple files to ensure tensors are initialized with zeros.
• Affected areas include model warmup, forward passes, grid and anchor grid initialization, stride computation, batch profiling, TensorBoard logging, and FLOPs calculation.

🎯 Purpose & Impact

• 🚀 Purpose: To avoid potential issues arising from using uninitialized memory by explicitly initializing tensors with zeros.
• 🔍 Impact: This change ensures consistency and stability during model inference and evaluation, making the system more robust and predictable.
• 🤖 For Users: Better reliability and determinism in model outputs, with less chance of erratic behavior due to uninitialized variables.

[pourmand1376]

I'm testing with colab to see if this one helps ...

[glenn-jocher]

Closing in preference of #9068 which isolates the issue better.