FIX: Preserve text columns when loading human-labeled CSVs

[Garyouki]

Description

Closes #2978.

HumanLabeledDataset.from_csv() loaded response 00123 as 123 and objective 00456 as 456 because pandas inferred numbers before the loader converted them back to strings. Read response, objective, harm-category, and data-type columns as strings in both the UTF-8 and Latin-1 paths. Human-label parsing and missing-value validation retain their existing behavior.

Tests and Documentation

• Added 16 CSV-loading cases spanning numeric/boolean-looking responses, objective/harm datasets, and UTF-8/Latin-1. Assertions cover original and converted response text, objective/category text, and typed human labels.
• Before the fix: 16 failed, 73 passed. After: 89 passed (pytest tests/unit/score/test_human_labeled_dataset.py -q --no-cov).
• Ruff lint/format, targeted ty check, async-suffix and reST-role checks, and git diff --check passed.
• No public API change. Full repository tests were not run.

[hannahwestra25]

thanks for contributing !