Normalize PromptIntel string filters before requests

[biefan]

Summary

• normalize validated severity and categories string filters during PromptIntel dataset initialization
• add regression coverage for string-based severity and category filters

Problem

_PromptIntelDataset validates string filter values such as severity="critical" and categories=["manipulation"], but it stores those values unchanged. _fetch_all_prompts() later assumes both fields are enums and accesses .value, which raises AttributeError before the request is sent.

This makes documented, validated string inputs fail at runtime instead of producing the expected API request.

Testing

• .venv/bin/pytest tests/unit/datasets/test_promptintel_dataset.py -q