Add Hyperbolic GCN layer

[julian-q]

Adds the HGCN layer from Hyperbolic Graph Convolutional Neural Networks to partially address #1334. This operator offers increased expressiveness over hierarchical graph topologies, e.g. tree structures. Adapted to PyG from the paper's original repo.

To benchmark performance against original implementation, you can clone my fork and follow the instructions in README.md. This implementation reproduces the paper's results, and is also quite a lot faster than the original due to vectorized manifold projections.

TODO

• [ ] Add full test coverage of:
  • [ ] HypLinear
  • [ ] HypAct
  • [ ] Hyperboloid
  • [ ] PoincareBall

[Padarn]

This looks really cool. Maybe an example of its use (to give people some idea of where it best applies) would also be nice?