Improve SimpleQA eval observability with structured logging and latency metadata

[SeanHe727]

Summary

This PR improves the SimpleQA evaluation pipeline by adding structured logging, per-query latency tracking, and reproducibility metadata.

Changes

• Added per-query latency measurement using time.perf_counter()
• Added structured JSON logging with:
  • run metadata: researcher model, timestamp, git commit
  • aggregate metrics: accuracy, average latency, p50 latency, p95 latency
  • per-query evaluation results
• Updated deprecated SMART_LLM_MODEL usage to SMART_LLM
• Added a guard for empty generator results to avoid next() failures
• Added documentation / example output for the structured eval log

Motivation

The previous eval output was primarily console-based, which made it harder to compare runs, track latency regressions, and reproduce evaluation results across model and configuration changes.

Compatibility

Existing eval behavior is preserved unless structured logging is enabled.

Testing

• Ran SimpleQA eval on a small sample
• Verified structured log output is valid JSON
• Verified aggregate latency metrics are generated correctly
• Verified existing console output still works

[SeanHe727]

Thanks for reviewing! This PR is intended as a small, backward-compatible improvement to make SimpleQA eval runs easier to reproduce and compare.

It adds structured JSON logging, per-query latency tracking, and run metadata without changing the default eval behavior. Happy to adjust the output format or scope based on maintainer preferences.