Add example for custom MPS kernels on macOS

Would be great to have a simple example showing how to build and use custom Metal (MPS) kernels with `torch.utils.cpp_extension` on Apple Silicon. Right now there’s nothing covering MPS, and it’s not obvious how to compile `.metal` files or register them with PyTorch. Even a minimal example (like an elementwise add) and a short note in the README would help a lot.
