Sequential CombinedLoader to flatten the eval and predict loops

[carmocca]

What does this PR do?

• Removes the _EvaluationEpochLoop
• Removes the _PredictionEpochLoop
• Removes the _DataLoaderLoop
• Moves the _EvaluationLoop out of the dataloader directory
• Moves the _PredictionLoop out of the dataloader directory
• Adds a loop.setup_data() method to the top-level loops. This method is called at the beginning of loop.run(). This replaces the trainer.reset_*_dataloader methods.
• Removes the trainer.reset_*_dataloader methods and related trainer.py simplifications
• The top-level loops now own their _DataLoaderSource and CombinedLoader
• The trainer.*_dataloader properties now return what the user returned in their Module.*_dataloader() hook. Nit: validation dataloaders are casted to list, see follow-up section
• Adds support for predict_step(dataloader_iter). this is not an important objective of the PR, but it's implemented for free.

Fixes #10696
Fixes #11845

Follow-ups:

• Refactor the setup_data implementations (Refactor loop.setup_data with utility functions #16918)
• Reduce the data connector (Refactor loop.setup_data with utility functions #16918)
• Move the num_*_batches ownership to the loops (Move max_batches definition to the Loops #16820)
• Trainer.evaluation_dataloaders should not be a forced list (Remove Trainer(multiple_trainloader_mode) in favor of CombinedLoader(mode) #16800)
• Allow passing CombinedLoaders, remove trainer flag for the mode (Remove Trainer(multiple_trainloader_mode) in favor of CombinedLoader(mode) #16800)
• Add a checkpoint migration for the loop state (Add a migration for the dataloader loops #17125)
• Review docs (Update data docs #16839)
• Move the training epoch loop outside of the epoch folder (Move _TrainingEpochLoop #16801)
• Move CombinedLoader file (Move the CombinedLoader to an utility file #16819)
• Check TODOs (Avoid instantiating CombinedDataset unnecessarily #16805, Add more pytree tests #16825, Consume the prediction batch indices iteratively #16826, Set a fixed stage in the evaluation loops #17094 )

cc @justusschock @awaelchli @Borda @carmocca

[justusschock]

That's a monster to review but seems pretty good. Only one question I might have missed due to the massive amount of changes I just tried to follow and a few mypy nits :)

A strong LGTM from me!

[awaelchli]

LGTM. This looks very promising.