Add capability for consistent GNN training from nekRS

[rickybalin]

This is a feature request to add the ability to train a consistent GNN online from nekRS data.

Consistency is currently enforced by running a pre-processing step (create_halo_info.py) on the nekRS data saved to disk in order to produce the needed data structures. This pre-processing step works in parallel, but does not scale very well.

As a first step:

• Add the computation of the consistency information to the Trainer class (in trainer.py) in order to generate the necessary data structures for the halo exchange and the consistent loss during online training.

As a second step:

• Look into scaling efficiency of the current parallel implementation of the consistency computation
• Perhaps augment the gnn plugin to provide information for consistency as well and make nekRS compute those data structures along with the other GNN and graph data.