Add skip_origin_metadata flag to optimize streaming startup time

[yuxin00j]

Fix #8197
This PR introduces the skip_origin_metadata kwarg to skip costly origin metadata fetching (ETags/mtime) when it's not needed, improving streaming initialization latency.

Changes

• Kwarg Plumbing: Routes the skip_origin_metadata kwarg from load_dataset(...) safely down to DatasetBuilder while stripping it securely from kwargs (via .pop()) so it's not leaked into underlying configuration classes that don't expect it (such as CsvConfig()).
• Streaming Validations:
  • Emits a warning when skip_origin_metadata=True is passed to load_dataset without streaming=True.
  • Disables download_and_prepare() entirely when builder._skip_origin_metadata is true by raising a ValueError("This function is not intended for streaming.").
• Test Coverage: Added robust test coverage spanning tests/test_data_files.py and tests/test_load.py:
  • test_DataFilesList_from_patterns_skip_origin_metadata
  • test_DataFilesList_from_patterns_skip_origin_metadata_raises_FileNotFoundError
  • test_load_dataset_builder_skip_origin_metadata (covers kwarg mapping & ValueError)
  • test_load_dataset_skip_origin_metadata_warning

[zhixiangli]

nit: why not skip_origin_metadata=config_kwargs.get("skip_origin_metadata", False)

[yuxin00j]

Fixed in the latest commit! I've inlined the .get() call here and then properly extracted it with .pop() when instantiating the builder later to prevent kwargs leakage.

[yuxin00j]

Hi @lhoestq, could you take a look at this PR? Thanks.