Fix dataset builder default version

[albertvillanova]

Currently, when using a custom config (subclass of BuilderConfig), default version set at the builder level is ignored: we must set default version in the custom config class.

However, when loading a dataset with config_kwargs (for a configuration not present in BUILDER_CONFIGS), the default version set in the custom config is ignored and "0.0.0" is used instead:

ds = load_dataset("wikipedia", language="co", date="20220501", beam_runner="DirectRunner")

generates the following config:

WikipediaConfig(name='20220501.co', version=0.0.0, data_dir=None, data_files=None, description='Wikipedia dataset for co, parsed from 20220501 dump.')

with version "0.0.0" instead of "2.0.0".

See as a counter-example, when the config is present in BUILDER_CONFIGS:

ds = load_dataset("wikipedia", "20220301.fr", beam_runner="DirectRunner")

generates the following config:

WikipediaConfig(name='20220301.fr', version=2.0.0, data_dir=None, data_files=None, description='Wikipedia dataset for fr, parsed from 20220301 dump.')

with correct version "2.0.0", as set in the custom config class.

The reason for this is that DatasetBuilder has a default VERSION ("0.0.0") that overwrites the default version set at the custom config class.

This PR:

• Removes the default VERSION at DatasetBuilder (set to None, so that the class attribute exists but it does not override the custom config default version).
• Note that the BuilderConfig class already sets a default version = "0.0.0"; no need to pass this from the builder.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[albertvillanova]

This PR requires one of these other PRs being merged first:

• Fix Version equality #4359
• Replace inequality with non-identity for inspect._empty doc-builder#211

[lhoestq]

Oh good catch ! Maybe it's worth adding a test to make sure a custom config gets the expected version ?

[lhoestq]

Thanks !