Implement synthetic tasks

[ekarais]

TODO:

• pass max_path_length to path_lifter.
• modify get_loaders() to select synthetic_chains.
• add optional sigmoid to model head depending on type of classification (?).
• create slurm script train_synthetic_chains.sh
• add equivariant update

[ekarais]

This is done. Not merging yet because we first need the following:

• separate wandb loggers for each task (we log the logits in the add-synthetic-tasks branch)
• feature normalization (this is done in the add-synthetic-tasks branch in an ad-hoc way by adding batch normalization before the message passing layer)