feat(automata): add ground-truth evaluation framework for classification accuracy

[AlexMikhalev]

Summary

• Generic evaluation framework for measuring automata classification accuracy against human-labeled ground truth
• Computes micro-averaged precision, recall, F1 overall and per-term
• Detects systematic errors (false positives appearing across multiple documents)
• load_ground_truth() for JSON-based ground truth files
• All types are serde-serializable for JSON report output

New types

GroundTruthDocument, ExpectedMatch, ClassificationMetrics, TermReport, EvaluationResult, SystematicError

Core API

pub fn evaluate(ground_truth: &[GroundTruthDocument], thesaurus: Thesaurus) -> EvaluationResult
pub fn load_ground_truth(path: &Path) -> Result<Vec<GroundTruthDocument>, std::io::Error>

Test plan

• 14 tests pass (all new, real thesaurus construction, no mocks)
• cargo clippy -p terraphim_automata clean
• cargo fmt clean
• CI green

Refs #576