Add a benchmark harness for the conflict classifier

[venkatapgummadi]

Background

We have unit tests for ConflictClassifier, but no benchmark to track classifier latency or accuracy
on a fixed corpus.

What needs to be done

Add ai-sync/benchmarks/classifier_benchmark.py that loads tests/fixtures/conflicts.jsonl, runs
classification, and prints accuracy + p50/p95 latency.

Definition of done

• Script runs with python -m ai_sync.benchmarks.classifier_benchmark
• Output saved as Markdown table
• Linked from docs/metrics-and-kpis.md
  Estimated effort: 3–4 hours · Skills needed: Python, basic ML evaluation