Fixes to parity inference task

[laserkelvin]

This PR provides some fixes to the ParityInferenceTask implemented in #294.

• Fixes incorrect mapping the in predictions property of ParityData, which caused targets and predictions to return the same tensor.
• Fixes missing entry in documentation for the general inference page
• Fixed mapping for collate_fn in the predict dataloader, in the case that only pred_split is specified and self.dataset is not defined as a Dataset object.
• QoL update to BaseInferenceTask class methods, allowing specific tasks to be loaded in, as well as a simplified pipeline to load in an artifact from wandb.