Static batch schema

[laserkelvin]

This PR addresses an issue with the existing BatchSchema definition, which is done on-the-fly based on what information was actually packed into the individual data samples. The issue with this is that serialization is not possible, which prevents multiple data loader workers being used, which hinders training.

This is solved by creating a static definition of a BatchSchema that subclasses DataSampleSchema. The scope of testing is on a new HDF5 dataset with PyG graphs, and is functional for multi-GPU training.

Summary of changes:

• Defined BatchSchema, subclassing DataSampleSchema
• Added a to method for DataSampleSchema (which is inherited by BatchSchema) for data transfers to accelerators
• Added transfer_batch_to_device methods to LightningModule definitions, which facilitates the data transfer before model calls
• Unit tests for batching with the new BatchSchema

[smiret-intel]

Does this imply we need to change the underlying data format to HDF5? I assume this works for native PyTorch as well.