feat: extend DeepSafe with scientific safety benchmarks

[coincidence2020]

Summary

This PR extends DeepSafe with three science-focused safety benchmarks while retaining the existing WMDP integration:

• SciHazard: evaluates harmful scientific responses using checklist matching, evidence search, executability, and novelty analysis.
• Safe-Scientist: evaluates whether models safely handle risky scientific research requests.
• SOSBench: measures over-safety and unnecessary refusal in scientific scenarios.

It also adds the corresponding datasets, evaluators, metrics, YAML configurations, and local execution scripts.

Implementation details

• Integrates the SciHazard DeHarm evaluation pipeline.
• Includes the SciHazard inputs and reproducibility-related checklist/search caches.
• Excludes prompt traces and completed experimental outputs.
• Keeps API credentials outside configuration files and loads them through environment variables.
• Updates both README.md and README_ch.md to document the four scientific safety benchmarks.
• Retains the existing WMDP implementation without modifying its original workflow.

Dataset availability

SciHazard inputs are included in this PR. Safe-Scientist and SOSBench dataset payloads are not redistributed; their expected local paths
are documented in docs/benchmark-data.md.

Verification

• 22 passed
• Checked for exposed credentials, internal service addresses, unpublished experiment commands, prompt traces, and unintended dataset
  outputs.