feat: eliminate requirement for custom root schema

[HennerM]

Currently each project needs to specify a root config schema and override the schemas of components that have a custom configuration, for example when having datasets that require a custom config. Additionally for callbacks and normalizers the way to add a Callback is cumbersome because we

1. create the callback class
2. create the callback config class
3. override the custom Trainer schema to allow the type of the new callback
4. use the callback in the config yaml

The 3rd step is just extra plumbing that can be avoided.

This PR adds functionality to avoid having to create a custom root schema. It works by designating certain component configs as "Registries". Allowing all subclasses of these registries to be valid types for validation, chosen by the already used "kind" discriminator. The semantics are the same to before, we just save a few things to write.

[github-actions]

Tests	Skipped	Failures	Errors	Time
1182	21 💤	0 ❌	0 🔥	26.443s ⏱️

[MauritsBleeker]

I would also remove the root config for the tutorial to be consistent (and change the text accordingly if needed)

[MauritsBleeker]

Why do we remove the type hinting here?

[HennerM]

We can't import BaseTrainer normally because that creates a cyclic import. That's why we did a if TYPE_CHECKING import on the BaseTrainer. However, with that it's not actually imported at runtime. This leads to an error when reading the type hints from the init init_params = get_type_hints(cls_.__init__), which is done to find the schema for a given class.

I think the if TYPE_CHECKING is kind-of bad anyway. We could fix the cyclic import by moving the imports in the trainer/base.py file inside the methods. I did that before and that ended up in a handful of local imports. I figured it's easier to just remove the import in this file instead, since we don't really need the type-hint for trainer here anyway