Added the conversion of scipy sparse to torch sparse. Wait until pyg …

[DomInvivo]

Wait for Pyg >= 2.4.0 being released released.

Checklist:

• Simplified the logic of the featurizer. Instead of matching the ordering between the adjacency matrix and the bonds, the adjacency is not generated directly from the bond orders.
• Improved unit-tests to verify that the ordering of the node, edge features and bond ordering is correct.
• Don't wait for the PR on torch_geometric since torch-sparse is too slow anyways, with limited memory benefits.
• Added a news entry: copy news/TEMPLATE.rst to news/my-feature-or-branch.rst) and edit it.

---

[DomInvivo]

@callumm-graphcore @joao-alex-cunha this PR changes the featurizer to return torch.Tensor instead of scipy.sparse.coo_matrix. I tried using torch.sparse_coo_tensor, but it made the featurization 5x slower on 32 CPUs.

This might help the dataloader, since it won't have to convert any numpy -> torch, nor any sparse -> dense.
Let me know if you see any speed/memory difference.

[DomInvivo]

The new version uses torch.Tensor instead of coo_matrix to store the features. The dataloading is ~10% faster since there's less conversion to do. I also tried torch.sparse_coo_tensor, but it was extremely slow.

In the image below, we see the difference between Tensor, coo_matrix, and Tensors while batching everything in a single graph with from_data_list.