Add benchmark reproduction command

[josephsenior]

Context

The README references a 4h33 autonomous run with 16,393 events and 39/39 passing tests,
but there's no easy way for others to reproduce a benchmark run on their own machine.

Proposed solution

• Add a grinta benchmark command (or a script in benchmarks/) that:
  • Clones a small, deterministic test repository
  • Runs Grinta with a predefined task prompt
  • Measures wall-clock time, token usage, event count, and test pass rate
  • Outputs a structured JSON report
• Include a benchmarks/README.md explaining how to run and interpret results
• Document expected baseline numbers for at least one model (e.g., GPT-4o, Claude Sonnet)

Why this matters

Reproducible benchmarks build trust and let contributors measure the impact of their changes.