feat(benchmark): add cross-framework A/B benchmark mode - #14

What

Implements G3 — cross-framework A/B benchmark mode: run_benchmark([BenchmarkCase(name, agent_a, agent_b)]) compares two agents head-to-head on the same task, scoring each case on four deterministic dimensions (steps, WEI, tokens, latency), majority-wins with explicit ties. Renders via summary() (CI logs) or to_markdown() (paste-ready table with overall tally).

Why

Roadmap WS-G3. With direct CrewAI/LangGraph ingestion landed (#12/#13), comparing a CrewAI run against a LangGraph run of the same task is now a natural ask — regression diffing answers "did we break?", A/B answers "which agent runs this leaner?".

How

• Both sides are peers: each case loads both traces, diffs them for TDI/structure context (compare()), and separately scores per-side efficiency dimensions from the traces themselves.
• Winner = majority of decided dimensions; equal values never count as wins; ties are first-class outcomes.
• Deliberately structural-efficiency only — no semantic judgment, honoring the LLM-judge non-goal (documented in module docstring).
• Mirrors the suite runner's execution contract: opt-in workers=N thread pool, input-order results, load/compare failures become contained ERROR outcomes.

Testing

• 9 new tests: identical-agent tie, majority verdicts both directions, partial dimension splits, error containment, summary/markdown content incl. exact table row + tally, parallel==sequential equivalence with order preservation, empty/invalid workers, and a true cross-framework case diffing the real CrewAI fixture against the real LangGraph fixture.
• Full suite: 300 passed; lint clean.

Checklist

• Exactly one logical change in this PR
• make lint passes
• uv run python -m pytest is green (300 passed)
• Docs updated if user-facing (public API docstring)
• CHANGELOG.md entry added under [Unreleased]
• Conventional title