Add ROCm/HIP support for AMD GPUs

[Hrzwahusa]

🚀 Add ROCm/HIP Support for AMD GPUs

📋 Summary

This PR adds full AMD GPU support to QPyTorch through ROCm/HIP, enabling low-precision arithmetic simulation on AMD hardware alongside the existing NVIDIA CUDA support.

🔧 Changes

• HIP Kernel Implementation: Ported all CUDA quantization kernels to HIP for AMD GPU compatibility
• Automatic Platform Detection: Build system now automatically detects and builds for either CUDA or ROCm/HIP
• Seamless Integration: No changes required to existing user code - works with the same API
• Comprehensive Testing: All existing tests pass on both CPU and AMD GPU (ROCm 7.1)

💻 Technical Details

• Added qtorch/quant/quant_hip/ directory with HIP kernel implementations
• Modified qtorch/quant/quant_function.py to detect ROCm and load appropriate extensions
• Updated .gitignore to exclude HIP build artifacts
• Tested on ROCm 7.1.25424

✅ Testing

Included test script (run_all_tests.sh) with logging. All tests pass:

• ✅ test_backward
• ✅ test_clamp
• ✅ test_device
• ✅ test_random
• ✅ test_relation
• ✅ test_stochastic

🎯 Benefits

• Enables QPyTorch on AMD GPUs (Tested on 7900XTX, should work with other models too)
• Maintains full backward compatibility with CUDA
• No performance degradation on existing CUDA setups

📦 Requirements

• ROCm >= 5.0 (tested with 7.1)
• PyTorch compiled with ROCm support