fix(multi-agent): preserve rollout logprobs

[Jiang020609]

Summary

• preserve token-level rollout logprobs in the multi-agent example when SGLang returns output_token_logprobs
• enable --use-rollout-logprobs in the multi-agent GRPO script so training can consume the saved rollout logprobs

Fixes #1976

Validation

• python -m ruff check examples/multi_agent/agent_system.py
• pre-commit run --files examples/multi_agent/agent_system.py examples/multi_agent/run-qwen3-30B-A3B-multi-agent.sh