Parity inference task

[laserkelvin]

This PR implements a ParityInferenceTask (with the lack of a better naming for it), which is used to run a datasplit through a model to evaluate predicted vs. actual, beyond just looking at reduced metrics like MSE/MAE. This task should be pretty straightforward to set up and run, and does not need additional logger configuration. After running trainer.predict, an inference_data.json is produced in the experiment folder (trainer.log_dir/<name>/inference_data.json) that can then be reviewed offline.

• Implements a ParityData structure which is mostly just a helper class for accumulating data.
• Implements the ParityInferenceTask, which in itself just provides the predict_step and on_predict_epoch_end functions to be called in the PyTorch Lightning predict pipeline.
• Unit test that verifies functionality on three datasets using EGNN. This tests simple functionality end-to-end, up to checking that the contents of the output JSON exists at all.
• Updated the documentation to include an "Inference" section, of which the first entry is this approach.

[melo-gonzo]

Looks good and useful! A lot of failing tests, not sure what is relevant. Double check if these are applicable, then good to merge

[laserkelvin]

I'm pretty certain the tests that fail are the usual suspects. I did pick up something that broke everything (fixed in 2faa68d), but otherwise seems to be fine. I'm going to merge!