Add SuperGPQA benchmark

[NotXf1le]

What changed

• Added a reproducible SuperGPQA benchmark runner for OpenRouter, llama.cpp, and Jev.
• Added deterministic dataset preparation and a 1,000-question stratified evaluation sample.
• Kept the evaluation sample separate from the 100-question pilot used to select working model and provider pairs.
• Added checkpoint/resume support, dataset integrity checks, pinned OpenRouter providers, cost tracking, and latency measurements.
• Added the benchmark chart and exact results to the README.
• Added tests for dataset preparation, deterministic sampling, pilot separation, and the pilot-size limit.
• Cleaned up benchmark names and reorganized the README reference sections.

Why

This provides a reproducible comparison of decision accuracy, cost, and latency across compatible language models and Jev. The dataset is downloaded separately and is not included in the repository.

Checks

• npm run check
• npm pack --dry-run