fix(scorers): reject non-finite normalized scores

[knisar]

Problem

ScoreNormalizer.from_scale() currently allows non-finite values. In particular, NaN reaches the clamping logic and becomes 1.0, so malformed judge output can turn into a passing score.

Scope

Validate numeric inputs at the shared normalization boundary.

Acceptance criteria

• Reject NaN, positive infinity, and negative infinity.
• Reject a non-finite max_value.
• Preserve the existing clamping and inversion behavior for finite values.
• A mocked judge response of "NaN" cannot produce an assessed pass through the core pipeline.

Non-goals

• Redesigning every judge response schema.
• Changing finite out-of-range clamping.
• Changing citation parsing in feat(scorers): add citation correctness scorer for RAG outputs #27 or retrieval parsing in feat: add RetrievalRelevanceScorer for RAG retrieval quality evaluation #37.

Likely files

• rai_toolkit/scorers/normalizer.py
• tests/test_score_normalizer.py

Testing

Tests must be deterministic and make no API calls.

Starting work

Before starting, check this issue's Development section and the open pull requests for linked work. Follow CONTRIBUTING.md by opening a draft pull request with Closes #38, your proposed validation behavior, and a short test plan. Keep the change limited to normalization and its regression tests.