Use diagonal batching in TorchDataset

[ArnNag]

Description

Currently, TorchDataset reads from an npz file that contains arrays of shape (n_conformers, n_atoms, *) for atom features. The atom feature array is padded to a constant n_atoms. This PR will revise HDF5Dataset to instead store an atom feature array of shape (n_atoms, *) and an array that tracks the number of atoms per conformer in an array of shape (n_conformers, ). TorchDataset.__getitem__ will still return an array of atom features for one conformer of shape (n_atoms, *), but it will do so by computing index ranges into the (n_atoms, *)atom feature array rather than direct indexing. Conformer features will still be stored in arrays of shape (n_conformers, *) as before.

Todos

• [x] Implement for new TorchDataset code in #11
• [x] Remove all hardcoded references to order of properties in properties_of_interest. Use format instead.

Status

• [x] Ready to go

[ArnNag]

The failing SchNet tests are because diagonal batching is not implemented within SchNet. test_dataset.py needs a test for a dataset with an arbitrary number of conformers in each record (e.g. SPICE). The existing tests were passing with incorrect implementations due to QM9 having one conformer per record.