Fix DatasetInfosDict metadata overwrite

[r-abaryan]

Fix DatasetInfosDict metadata overwrite

Summary

DatasetInfosDict.write_to_directory(..., overwrite=True) raises
AttributeError: 'dict' object has no attribute 'to_dataset_card_data' for
nonempty input. The writer initializes its collection as a plain dict; if a
legacy dataset_infos.json exists, it can update that file before crashing and
leave the README metadata stale.

Initialize a DatasetInfosDict and replace the README's existing
dataset_info field when overwriting. An empty overwrite now clears old
configuration metadata while preserving other card metadata and description.

Validation

• Added regression coverage for fresh directories, README and legacy JSON
  layouts, new/existing/empty configurations, merge and overwrite behavior,
  multiple configurations, and preservation of unrelated card content. The
  new tests exposed 11 failures before the fix.
• tests/test_info.py tests/test_metadata_util.py: 50 passed on both Python
  3.10.21 and 3.14.7.
• Windows Python 3.14 unit suite: 3,318 passed, 232 skipped, with three tests
  requiring unavailable symlink privileges excluded.
• Repository-wide Ruff lint and formatting checks passed.

The broader Windows run excluding CI Hub tests completed with 3,399 passed,
258 skipped, 4 integration failures, and 3 symlink setup errors. Those failures
did not enter the changed writer; they involved missing torchcodec, a
subprocess using base Python, an old cache hash expectation, and Windows
symlink privileges. The CI Hub module stalled on a cleanup request without a
timeout, so a clean full-suite result is not available locally.