Simplify `PCDLit` and `PCDDataModule` with `hparams`

[adosar]

Thanks to self.save_hyperparameters() all init argument are available via self.hparams or the immutable self.hparams_initial. This means it is not necessary to use dataclass for reducing boilerplate.

Also, consider removing writing to hparams.yaml since this can take a long time if the model is large. See also:

• Do not save the weights of the model to hparams.yaml Lightning-AI/pytorch-lightning#11494 (comment)
• https://lightning.ai/docs/pytorch/stable/common/lightning_module.html#id2
• Incorrect .hparams when using LightningCLI with Callable Lightning-AI/pytorch-lightning#20726

Note

Care must be taken when using the LightningCLI because it seems that self.hparams will store the (un)instantiated object as dict.

[adosar]

In PCDLit there are no parameters that can be tracked in the hparams tab (dict's like config_optimizer can't be shown there). On the other hand, in PCDDataModule there are parameters that can be tracked (e.g. train_batch_size). I don't have a strong opinion here, but at least calling self.save_hyperparameters(logger=False) won't save weights in the hparams.yaml.

Regarding using hparams, see here first: Lightning-AI/pytorch-lightning#20777 before applying any change.