feat(benchmarks): add reproducible generation benchmark

[vetertann]

Linked Issue

Closes #207
Supersedes #239

Description

This PR ports the structured generation benchmark to TypeScript and integrates it into the existing benchmarks workspace. It ships the reproducible harness, gold fixtures, and the original 21-model result set together.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Performance improvement
• Test coverage improvement

Changes Made

• Added a TypeScript benchmark runner for plain JSON, JSON object mode, and TOON generation.
• Added strict schemas, manual validation, canonicalization, and up to two repair attempts.
• Uses the local TypeScript TOON encoder/decoder without Python or CLI subprocesses.
• Added fixture generation and committed JSON/TOON gold fixtures for all four cases.
• Added the historical 21-model raw and aggregate result CSVs.
• Added Nebius configuration, model/run selection, checkpoints, request timeouts, and usage accounting.
• Documented setup, methodology, commands, and the published baseline results.

SPEC Compliance

• This PR implements/fixes spec compliance
• Spec section(s) affected: N/A (benchmark tooling)
• Spec version: N/A

Testing

• All existing tests pass
• Added new tests for validation, canonicalization, wrapper extraction, and result aggregation
• TypeScript typecheck passes
• ESLint passes
• Full suite passes (644 tests)
• End-to-end Nebius smoke run with Qwen/Qwen3.5-397B-A17B

Pre-submission Checklist

• My code follows the project style guidelines
• I have performed a self-review of my code
• I have added tests that prove the change is effective
• New and existing tests pass locally
• I have updated the documentation where needed
• My changes generate no new warnings

Breaking Changes

• No breaking changes

Additional Context

The committed results preserve the published baseline: 10 runs per case and model across 21 models via Nebius. Future runs can select models and run counts through environment variables.