feat: machine-readable eval report for LLM optimization agents

[StephaneWamba]

This builds on #111 (output hash / correctness check) to close the loop for automated optimization agents.

The problem with #111 alone: the agent gets a human-readable terminal output but still needs to parse unstructured text and know the Python API to make a binary accept/reject decision.

What this adds:

--codspeed-eval-report=eval.json -- write a machine-readable JSON file after each comparison run:

{
  "aggregate_score": 0.33,
  "is_acceptable": true,
  "benchmarks": [
    {
      "name": "tests/test_sort.py::test_sort",
      "perf_gain": 0.33,
      "output_changed": false,
      "score": 0.33
    }
  ]
}

The agent reads this file and makes a decision. No Python knowledge needed, works from any language or shell script.

Scoring logic (also in #111):

• aggregate_score is the minimum score across all benchmarks -- conservative by design, one correctness failure vetoes the whole suggestion
• score = perf_gain when output is correct, 0.0 when broken, null when unknown (no --codspeed-capture-output)
• is_acceptable = aggregate_score > 0.0 and not nan

Demo:

examples/optimize_loop.py shows the full loop: baseline run -> apply patch -> rerun with --codspeed-eval-report -> read JSON -> print verdict. Runs standalone with python examples/optimize_loop.py.

Tests added:

• test_eval_report_written_after_second_run -- file is created, has the right keys
• test_eval_report_acceptable_when_output_stable -- output_changed: false for unchanged output
• test_eval_report_not_acceptable_when_output_breaks -- aggregate_score: 0.0, is_acceptable: false
• test_eval_report_not_written_on_first_run -- no baseline means no file
• Unit tests for aggregate_score, is_acceptable, to_dict including nan->null serialization

Depends on #111.