Add skip_origin_metadata flag to optimize streaming startup time

[yuxin00j]

Resolves #8197.

When streaming datasets that consist of many remote data files, the initial origin metadata fetch (_get_origin_metadata which queries ETag/mtime) can become a significant latency bottleneck.

This PR implements the accepted feature request to introduce an optional skip_origin_metadata keyword argument to bypass these checks and improve streaming startup time.

Changes:

• Added skip_origin_metadata: bool = False to DatasetBuilder.__init__ and data resolution methods across DataFilesList and DataFilesDict.
• When skip_origin_metadata=True, pattern resolution assigns empty tuple placeholders [()] * len(data_files) rather than making expensive network/filesystem calls.
• The standard file-existence check is preserved during resolution; if a file doesn't exist, it correctly raises a FileNotFoundError.
• Added validation in load_dataset: this optimization is strictly intended for streaming mode. If skip_origin_metadata=True is passed alongside
  streaming=False, a warning is logged and the flag is safely forced back to False.
• Added tests to tests/test_data_files.py and tests/test_load.py to ensure correct behavior and error handling.