docs: add LLM-to-LLM conversation eval example

[deepujain]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor

What is the current behavior? (You can also link to an open issue here)

Fixes #2803.

The multi-agent docs describe handoffs, tools, and explicit agent workflows, but they do not include a compact example for fixed-turn LLM-to-LLM evaluations where two agents converse and a third model judges the transcript. Users trying to build client-agent/customer-service-agent style evals have to infer the pattern from lower-level agent APIs.

What is the new behavior?

The multi-agent docs now include an LLM-to-LLM conversation example that:

• Alternates a client agent and service agent through an explicit solver workflow.
• Uses AgentState and run() so the two agents share and extend the same conversation history.
• Adds a third judge model scorer over the transcript.
• Notes that submit=False is useful when the outer workflow controls turn-taking, and points users to message/token/time limits for tool-using agents that may run too long.

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

No. This is a docs-only addition for existing agent and scorer APIs.

Other information:

Validation:

• Targeted:
  • uv run pytest tests/agent/test_agent_execute.py -v
    • 31 passed, 6 skipped locally.
    • Confirms the run() execution path used by the example.
    • Confirms agent invocation as solver/tool/handoff and limit handling, which supports the guidance about bounded agent runs.
  • Reviewed the docs example against the current react, run, AgentState, solver, and scorer APIs after rebasing onto current origin/main.
• Regression:
  • uv run make check
  • uv run make test
  • Commit-time pre-commit hooks passed: large-file, JSON/YAML, debug-statement, private-key, requirements, and docs typo checks.

CI/CD coverage expected:

• Standard Build workflow should cover ruff, ruff format, mypy, pre-commit, package build, and pytest on Python 3.10 and 3.11.
• Docs are excluded from ruff's source lint, but the pre-commit typos hook covers markdown/docs text.
• Log viewer and sandbox-tools special gates should not require extra artifacts because this PR only changes multi-agent docs.

Closes #2803.