Fix HASH vector fetch decoding and score normalization

[sagile]

Summary

• decode HASH-stored vector fields on fetch so callers get logical number[] values instead of raw binary-looking payloads
• normalize vector search scores using the query alias and schema metric so normalizeDistance: true returns similarity-style scores consistently
• add unit and integration coverage for both the fetch round-trip and the manual smoke normalization flow

Validation

• npm run format:check
• npm run lint
• npm run type-check
• npm run build
• npm run test -- tests/unit/indexes/search-index.test.ts
• npm run test -- tests/integration/search-index.test.ts

CI Notes

• Reviewed .github/workflows/lint.yml and ran the local equivalents: npm run format:check, npm run lint, and npm run type-check
• Reviewed .github/workflows/test.yml and ran focused local equivalents for the changed surface: npm run build, npm run test -- tests/unit/indexes/search-index.test.ts, and npm run test -- tests/integration/search-index.test.ts
• The PR workflow also runs broader test-matrix jobs (tests/unit, tests/integration, full npm test) on Node 22/24 and Redis variants; those were not fully re-run locally because the touched surface is narrower than the full matrix