v0.2.5

This release contains the following changes:

• Updating the O3MessagePassingBlock used by MACE and NequIP to use e3j.core.Convolution, which enables fused convolution kernels on GPU and TPU but requires an explicit edge ordering to be set.
• Adding the Graph.ordering field to specify the edge ordering explicitly, and the Graph.sort_edges() helper. All graphs created within the library are now sender-ordered by default, matching the default setting in O3MessagePassingBlock.
• Changing the default target_irreps in NequipConfig to use 32 channels for l=2 features (previously 4), increasing the default model size.
• Adding num_reader_workers to GraphDatasetBuilderConfig to enable reading data using a worker pool. Used to read multiple datasets or splits in parallel, and to also parallelize over groups inside HDF5 files.
• Changing the default setting of Graph.edges.shifts from all-zeros to None when a system does not have PBCs. This can greatly reduce the memory required to hold a processed dataset in RAM during training.
• Fixing a bug where the property_name_mapping was not correctly handled in the ASEAtomsReader.
• Reducing the runtime of the GraphDataset preparation workflow by reducing device-to-host memory transfers when shuffling, removing unnecessary iterables over the batched dataset, and simplifying the computation of per-element atomic energies.
• Filtering out graphs with no edges during dataset building, and guarding the minimum-neighbor-distance statistic against edgeless graphs.
• Adding a terminate_on_nan option to TrainingLoopConfig that stops training as soon as an epoch's mean loss or gradient norm becomes NaN/Inf.
• Upgrading e3j dependency to minimum version of 0.1.1.