Node embedding projections for ForceRegressionTask

[laserkelvin]

This PR refactors ForceRegressionTask to regress energies by reducing node contributions, as opposed to system-level embedding regression.

• Energy output head sees node embeddings now, rather than graph embeddings
• Define readout functions that depend on whether a graph structure is passed, and what framework is used. The reduction can be configured with the task hyperparameter embedding_reduction_type
• process_embeddings will pack the node energy contributions into the output dictionary.

[laserkelvin]

The force task tests pass

[melo-gonzo]

LGTM! Just out of curiosity, but what would the use case be for using a readout other than 'sum' for a graph level energy?

[laserkelvin]

You can also do a mean reduction, or if you want to be super fancy, an attention weighted one (like in graph attention). Sum is more physically meaningful than a mean reduction IMO