Overhaul deprecations system in dask.config

[crusaderky]

The config deprecation system was problematic:

• dask/dask used check_deprecations, which worked for dask.config.set but not for ~/.config/dask.yaml
• dask/distributed used rename, which worked for ~/.config/dask.yaml but not for dask.config.set

This PR overhauls and unifies the whole system.
Additionally, it adds support in rename for dot-separated old keys and None values.

Follow-up: dask/distributed#8179

[crusaderky]

Deliberately not linking this to the sphinx docs

[jrbourbeau]

We have a separate check_deprecations method for handling deprecations here -- does this result in multiple warnings being emitted?

[crusaderky]

Ah. I completely missed that.
This looks like a bit of a mess. dask/dask uses check_deprecations, while dask/distributed uses rename. The two systems are completely independent.

I'll rework this PR to unify them somehow.

[crusaderky]

Upon further inspection:

• dask/dask uses check_deprecations, which works for dask.config.set but not for ~/.config/dask.yaml
• dask/distributed uses rename, which works for ~/.config/dask.yaml but not for dask.config.set

I overhauled and unified the whole system.
Follow-up: dask/distributed#8179

[crusaderky]

@jrbourbeau this PR is ready for review again

[crusaderky]

@jrbourbeau this is now blocking #10507

[jrbourbeau]

Thanks @crusaderky

[bnaul]

I think this may have introduced an accidental change in the behavior of dask.config.get (I think dask/dask-kubernetes#816 is referring to the same thing, albeit somewhat cryptically 🙂):

# Previous behavior, 2023.9.1 and earlier
HEAD is now at 3316c38f0 bump version to 2023.9.1
(.venv)  ➜  dask git:(2023.9.1) ✗ python -c 'import dask; print(dask.config.get("array.chunk-size", override_with=None))'
128MiB

# New version after this PR
HEAD is now at 216c7c939 Release 2023.9.2 (#10514)
(.venv)  ➜  dask git:(2023.9.2) ✗ python -c 'import dask; print(dask.config.get("array.chunk-size", override_with=None))'
None

I believe this also contradicts the docstring so I think it's a clear bug and not just a neutral change

If override_with is not None this value will be passed straight back

cc @crusaderky @jrbourbeau @fjetter

edit: opened #10519 to discuss