pyg-team/pytorch_geometric ? reverse-engineered prompt

Build me a Python library for graph neural networks on top of PyTorch, something that makes it easy to train models on graph data, point clouds, and other irregular structures.

I want a clean, friendly API that feels close to normal PyTorch, with a few built in graph layers like a basic graph convolution, a message passing base class, dataset loaders, and simple helpers for batching lots of small graphs or one big graph. It should also include a couple of example training scripts, like node classification on a citation network, plus enough tests and docs so someone can actually learn how to use it.

If it helps, look up current docs online for any tricky PyTorch details. Keep the code organized and practical, with support for common benchmark datasets, custom transforms, and room to extend the library with new layers later.