fix: preserve full Groq response schemas in evals

[cwilmott-groq]

Summary

• retain the original JSON Schema alongside Inspect typed schemas so Groq receives oneOf, conditionals, patterns, lengths, and other unmodeled assertions
• validate each schema with its declared dialect instead of forcing every input through Draft 2020-12
• normalize nullable type arrays for the Inspect compatibility representation
• omit null response-schema descriptions from Groq requests
• validate decimal multipleOf exactly and use RFC3339 time semantics in the scorer

Why

Inspect previously discarded assertions before inference while the scorer retained them, creating false constrained-decoding failures. This change makes the sent schema and scored schema agree.

Validation

• 15 targeted provider, scorer, and JSONSchemaBench dataset tests pass
• Ruff formatting and lint checks pass