Add example and documentation for linking external NVIDIA libraries

[0x00b1]

The current examples in extension-cpp focus on simple CUDA kernels or extensions that rely only on libraries already linked by PyTorch. It would be very helpful to include an example and documentation demonstrating how to link additional NVIDIA libraries that PyTorch does not automatically link (e.g., AmgX, cuSOLVER, OptiX).

This would clarify:

• How to specify additional -L and -l flags in torch.utils.cpp_extension.load or setup.py-based builds
• How to manage include paths (include_dirs, library_dirs) and potential conflicts with PyTorch’s existing CUDA setup
• Platform-specific nuances