Normalize LLM judge scores in result JSON

[Sehlani042]

Summary

• normalize raw LLM judge subdimension scores from the 0-10 rubric to the 0-1 result scale
• recalculate aggregate judge scores from normalized subdimensions for both standard and stability evaluation paths
• document the score scale and add unittest coverage for helper, stability averaging, and TaskEvaluator output

Tests

• python3 -m unittest discover -s tests
• python3 -m compileall benchmark tests/test_score_normalization.py
• git diff --check

Fixes #15.
Addresses #23.