[codex] collect server metrics during benchmarks

[jamesdborin]

Summary

• add JSONL dataset support and request model injection for CLI benchmarks
• ignore experiments/ so local benchmark artifacts do not get staged accidentally
• add server-side Prometheus metrics scraping during benchmark runs

Metrics details

• adds --metrics-output-dir, --metrics-endpoint, --metrics-interval-ms, --metrics-timeout-ms, and --metrics-fail-on-error
• captures a baseline scrape before workers start and writes raw scrapes, parsed samples, metadata, and a derived summary
• exposes metrics scrape metadata through BenchmarkReport, Python bindings, and CSV output
• updates the Prime harness to copy remote metrics artifacts back into the local run directory

Validation

• cargo test --manifest-path rust/Cargo.toml
• python3 -m py_compile src/batchbench/harness.py
• local smoke test with mock /v1/chat/completions and /metrics endpoint