docs: fix factual errors in multi-agent example page

[Shi-Dong]

Summary

Fixes factual errors in docs/examples/multi-agent.md found during a fact-check of every page under docs/examples/ against the actual code in examples/multi_agent/.

• Pipeline description: the page described a "dual-agent thinker/verifier" system with alternating turns. The shipped example is actually a three-role pipeline — parallel solvers, a round of rewriters that revise the solvers' answers, and a selector that picks the best one (agent_system.py: run_agent_system → solver_worker / rewrite_worker / SelectorAgent). Intro, walkthrough, and variations rewritten to match.
• Configuration/tuning: removed the nonexistent MAX_TURNS knob (MULTI_AGENT_CONFIGS has no such key); num_parallel is the number of concurrent solver/rewriter attempts per prompt, not a turn count.
• What to watch: the doc listed fabricated multi_agent/* and loss_mask/role_split metrics. Replaced with the real ones (rollout/raw_reward, perf/rollout_time, standard step-line metrics) plus how rewards are scaled by correct_reward_weight / incorrect_reward_weight.
• Troubleshooting: rows referencing the nonexistent verifier/MAX_TURNS replaced with real failure modes (rewriter template strength in prompts.py, deepscaler RM needing a boxed answer, sequential solver→rewriter→selector round-trips).
• VLM variation: referenced a nonexistent call_role parameter; rewritten in terms of making the solver/rewriter workers VLM-aware.

Test plan

• Verified against examples/multi_agent/agent_system.py, prompts.py, and rollout_with_multi_agents.py on current main
• Docs render check

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.