Decouple Tuner from Trainer

[awaelchli]

What does this PR do?

Removes:

• Trainer.tuner
• Trainer.tune()
• Trainer(auto_lr_find=..., auto_scale_batch_size=...)

New usage:

trainer = Trainer()
tuner = Tuner(trainer)
tuner.lr_find(...)
tuner.scale_batch_size(...)

Furthermore, the same tuner features are available through the LearningRateFinder and BatchSizeFinder callbacks.

Note: The tuning callbacks still rely on the Trainer handling the TunerExitException. Haven't yet found a way to decouple that. This can be discussed in a follow-up.

Fixes #5374
Fixes #9103

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

Yes, justified by the 2.0 initiative of making Trainer leaner.

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?
• 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 update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• 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

Did you have fun?

I made sure I had fun coding 🙃

cc @Borda @justusschock

[carmocca]

Phew! Nice work

[Borda]

I love the simplification, just thinking about the new name...

[justusschock]

Implementation looks good. Just a general question:
Do we really need a new tuner object or can these be callbacks only?

[awaelchli]

Do we really need a new tuner object or can these be callbacks only?

The tuner object is not new, it was always a public class (maybe it should not have been from the start). The callbacks were introduced recently yes, but there are features in the tuner that are a bit unclear how you would use them with just callbacks. For example, an interaction like this:

tuner = Tuner()
finder = tuner.lr_find(...)
finder.plot()
finder.suggestion()

etc.

The good thing is, after this PR, the development on the tuner code can happen without impacting anything on the Trainer side. So there is still the possibility to go in either direction.