Do not prefetch when possible

[carmocca]

What does this PR do?

Only prefetch unless we require it. That is when:

• Iterable datasets are used, where we need to know when it will be exhausted to run validation.
• Interbatch parallelism is used, where no prefetching would be silly.

This is good for FFCV because they already do prefetching internally.

We no longer error on 0-length map-style datasets as this is inconsistent with iterable datasets.

Part of #11538
Follow-up to #11606
Reverts #1280

Does your PR introduce any breaking changes? If yes, please list them.

None knowingly

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• [n/a] Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

cc @Borda @justusschock @awaelchli @ninginthecloud @akihironitta

[tchaton]

LGTM ! Some comments.

[rohitgr7]

small comments.

[rohitgr7]

Suggested change

	when the training will stop and run validation if necessary.
	when the training epoch will end and run validation if necessary.

[carmocca]

It can happen at any point in training:

https://github.com/PyTorchLightning/pytorch-lightning/blob/02ccd874b9eda885df7752710e605e259d579d9c/pytorch_lightning/loops/epoch/training_epoch_loop.py#L488-L489

[awaelchli]

while the comment here is not wrong, the real reason we have the prefetching is actually to avoid starting a new epoch that doesn't have any batches left in the dataloader. Relying on a StopIteration check is not possible for this reason.