feat: add RiskAssessment.load() for cross-session reuse

[frankentini]

Summary

Adds a RiskAssessment.load(path) classmethod that deserializes JSON files produced by RiskAssessment.save(), enabling users to reload previous risk assessment results across Python sessions.

This addresses the core ask in #200 — save/load for iterative red-teaming workflows where you want to generate attacks once and re-run them against improved model versions.

What it does

• RiskAssessment.load(path) — reads a saved JSON file and reconstructs a fully typed RiskAssessment instance
• Enum restoration — builds a reverse lookup across all VulnerabilityType enum classes to map serialized string values (e.g. "gender") back to their proper enum members (e.g. BiasType.GENDER)
• Multi-turn support — reconstructs RTTurn objects for multi-turn test cases
• Graceful fallback — unknown/custom vulnerability type values are preserved as strings rather than raising

Usage

# Save after a red-teaming run
risk_assessment.save('./results/')

# Load in a new session
loaded = RiskAssessment.load('./results/20260402_081100.json')

# Re-inject into a RedTeamer for re-evaluation
red_teamer.test_cases = loaded.test_cases

Tests

Added comprehensive test suite covering:

• Roundtrip save → load with enum integrity
• Multi-turn test case reconstruction
• Multiple vulnerability types in one assessment
• Error handling (missing file, invalid JSON)
• Enum reverse-lookup correctness

Closes #200

[vercel]

@frankentini is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.