Read cached dataset_info.json to populate config_names offline

[adityasingh2400]

What

get_dataset_config_names with HF_DATASETS_OFFLINE=1 returns ['default'] even when the cache contains multiple config subdirectories for the dataset. For example, after caching cais/mmlu:

$ HF_DATASETS_OFFLINE=1 python -c "import datasets; print(datasets.get_dataset_config_names('cais/mmlu'))"
Using the latest cached version of the dataset since cais/mmlu couldn't be found on the Hugging Face Hub (offline mode is enabled).
['default']

This is because CachedDatasetModuleFactory.get_module returns a DatasetModule without builder_configs_parameters, so get_dataset_config_names falls back to default.

Fix

Walk the cached repo subdirectories, parse each dataset_info.json, and populate builder_configs_parameters with the discovered config names. Malformed or unreadable dataset_info.json files are skipped silently so a single bad cache entry does not break the lookup.

With this change:

$ HF_DATASETS_OFFLINE=1 python -c "import datasets; print(datasets.get_dataset_config_names('cais/mmlu'))"
Using the latest cached version of the dataset since cais/mmlu couldn't be found on the Hugging Face Hub (offline mode is enabled).
['abstract_algebra', 'all', 'anatomy', ...]

Test

Adds test_CachedDatasetModuleFactory_offline_populates_config_names in tests/test_load.py that builds a fake cache layout with three configs, calls the factory with HF_HUB_OFFLINE patched to True, and asserts every cached config name is returned. No network access required.

Fixes #7947. Supersedes the abandoned #7977.