[RFC] Remove has_{setup/teardown}_{stage} properties from DataModule

[ananthsub]

🚀 Feature

Carrying forward discussion from #7289 (comment)

Remove these settings to give more control to the users and making trainer execution easier
https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/core/datamodule.py#L33-L52

Motivation

Consecutive calls like this have different behavior:

dm = MyDataModule()
model = MyLightningModule()
trainer.fit(model, datamodule=dm)
trainer.fit(model, datamodule=dm)

The second trainer.fit call will not run the prepare data, setup, or teardown hooks.

Removing this check gives us these advantages

1. Simpler, consistent hook execution: More control for users for how they want to handle state changes
2. Less state management inside the trainer and datamodule: less framework overhead/maintenance for us long term and easier debuggability

cc @carmocca

Pitch

Remove this code:

• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/core/datamodule.py#L27-L97
• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/core/datamodule.py#L162-L173
• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/core/datamodule.py#L230-L309
• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/trainer/trainer.py#L1169-L1170
• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/trainer/connectors/data_connector.py#L63
• https://github.com/PyTorchLightning/pytorch-lightning/blob/490cc57809ebeba19003b4101393a8a058217c31/pytorch_lightning/trainer/trainer.py#L1143-L1144

cc @carmocca

Alternatives

Additional context

[kandluis]

+1 that this would be helpful in simplifying the overhead of the DataModule API.

[carmocca]

Continuing discussion from @tchaton #7238 (review):

I think this is an important feature as preparing a dataset can be costly and shouldn't be done several times.
How do you propose to solve this problem if we remove those properties ?

The argument is that the users should implement the checks for "should download/prepare" themselves, giving them more power and responsibility.