feat(cookbooks): add live CrewAI e2e recipe through openinference adapter - #5

What

Adds cookbooks/live_crewai.py: a live, end-to-end CrewAI example. It runs a real one-agent crew (custom BaseTool + gpt-4o-mini), instruments it with openinference-instrumentation-crewai + openinference-instrumentation-openai, and diffs two real crew executions through the existing openinference adapter — no manual trace reshaping.

Why

Roadmap item J3 — real end-to-end framework examples. CrewAI is the second most-adopted agent framework (5.2M downloads/mo per WS-A); this proves AgentDiff ingests its telemetry natively and gates regressions on real crew output (verified live).

How

• Baseline: single NY lookup. Candidate: NY+CA+TX scope creep → 3 tool iterations.
• Integration gotchas discovered and baked into the recipe:
  • Tool must be a BaseTool subclass overriding _run — @tool-decorated tools override run() and bypass the instrumentor's patch (no tool spans).
  • CrewAI 1.x dropped litellm; its LLM layer calls the OpenAI SDK directly, so OpenAIInstrumentor provides LLM spans.
  • CrewAI product telemetry / interactive trace prompt disabled (CREWAI_DISABLE_TELEMETRY, CREWAI_TRACING_ENABLED, tracing=False).
• Live verification (2026-08-22): TDI 0.429, loop flag across 3 LLM+tool iterations, tokens +218%, gate BLOCKED on divergence AND loops. Evidence recorded in context/LIVE-VERIFICATION.md.
• Updates cookbooks/README.md + CHANGELOG [Unreleased].

Testing

• Ran the recipe live against the real API: gate blocked as designed
• make lint -> pass; uv run python -m pytest -> 205 passed

Checklist

• Exactly one logical change in this PR
• make lint passes
• uv run python -m pytest is green (205 passed)
• Docs updated if user-facing
• CHANGELOG.md entry added under [Unreleased]
• Conventional title