feat(native): Add C++ CUDA backend with pybind11 bindings

[m96-chan]

Summary

• Add native C++ CUDA backend using CUDA Runtime/Driver API and NVRTC
• Implement pybind11 bindings to expose C++ functionality to Python
• Replace cuda-python dependency with native C++ module

Changes

New native/ directory

• core/: Device management, GPUArray, Stream using CUDA Runtime API
• jit/: NVRTC compiler and JITKernel for runtime CUDA kernel compilation
• ops/: CUDA kernels for add, mul, matmul operations
• bindings/: pybind11 bindings exposing C++ to Python
• CMakeLists.txt: CMake build configuration

Python layer updates

• backend.py: Replace CUDABackend with NativeBackend that loads native module
• compiler.py: Use native JITKernel for NVRTC compilation
• basic.py: Use native ops when native backend is available

Architecture

Python API (pygpukit)
    │
    ├── CPU Simulation Backend (testing/fallback)
    │
    └── Native Backend (pybind11)
            │
            └── C++ Module (_pygpukit_native)
                    │
                    ├── core/ - CUDA Runtime API
                    ├── jit/  - NVRTC
                    └── ops/  - CUDA Kernels

Test plan

• All 73 Python tests pass with CPU simulation backend
• Build native module with CMake + CUDA Toolkit
• Test on GPU hardware

Build Requirements

• CUDA Toolkit 11.0+
• CMake 3.18+
• pybind11

cd native
mkdir build && cd build
cmake ..
make

🤖 Generated with Claude Code