feat: add GPU support

[hspedro]

🚀 Add CUDA Support for GPU-Accelerated Translation

Summary

This PR adds proper CUDA support to the Babeltron translation service, enabling GPU-accelerated inference for both NLLB and M2M100 translation models. The changes ensure that PyTorch correctly detects and utilizes NVIDIA GPUs when available, significantly improving translation performance.

Changes

🔧 Docker Configuration

• Updated the Dockerfile to install CUDA-enabled PyTorch instead of the CPU-only version
• Added explicit installation of PyTorch with CUDA 11.8 support
• Removed unnecessary CUDA runtime dependencies that were causing build failures
• Maintained compatibility with CPU-only environments for development and testing

🧠 Model Optimization

• Verified that both NLLB and M2M100 models properly detect and utilize GPU acceleration
• Ensured proper fallback to CPU when GPU is not available
• Maintained the existing architecture detection logic for CUDA, MPS, ROCm, and CPU

📝 Documentation

• Added instructions for setting up the development environment with GPU support
• Updated deployment documentation with GPU requirements
• Added troubleshooting section for common GPU-related issues

Testing

• Verified CUDA detection with torch.cuda.is_available()
• Confirmed GPU acceleration works with both NLLB and M2M100 models
• Tested translation performance improvements (approximately 5-10x faster inference)
• Ensured backward compatibility with CPU-only environments

Dependencies

• Updated PyTorch to use CUDA 11.8-compatible version
• Maintained compatibility with existing transformers library version

Deployment Notes

To deploy this version with GPU support:

1. Ensure the host has NVIDIA drivers installed
2. Install the NVIDIA Container Toolkit (nvidia-docker2)
3. Use the updated docker-compose.yml which includes GPU device mapping

---

Note: This PR requires a host with NVIDIA GPU and properly configured drivers to fully utilize the GPU acceleration features. The application will still function on CPU-only environments but with reduced performance.