feat: add AG2 multi-agent framework as native agent client

[faridun-ag2]

Summary

Adds AG2 (formerly AutoGen) as a native agent client for AgentBench evaluation. AG2 is an open-source multi-agent framework with 500K+ monthly PyPI downloads, 4,300+ GitHub stars, and 400+ contributors.

• AG2Agent: Single AssistantAgent + UserProxyAgent pair (comparable to existing HTTPAgent)
• AG2GroupChatAgent: Multi-agent GroupChat with Planner + Executor (demonstrates AG2's multi-agent advantage)

Dependency strategy

AG2 v0.11.4+ requires Python >=3.10 and pydantic >=2.6.1, which conflicts with AgentBench's core dependencies (Python 3.9, pydantic ~=1.10). Rather than modifying existing requirements, a separate requirements-ag2.txt is provided for an isolated venv:

python3.10 -m venv .venv-ag2
source .venv-ag2/bin/activate
pip install -r requirements-ag2.txt

The AG2 import in __init__.py is wrapped in try/except ImportError so the base AgentBench installation is unaffected.

New files

• src/client/agents/ag2_agent.py — AG2Agent and AG2GroupChatAgent implementations
• src/client/agents/test_ag2_agent.py — 23 tests (unit, integration, e2e with fake task server)
• configs/agents/ag2-single.yaml — Config for single-agent evaluation
• configs/agents/ag2-groupchat.yaml — Config for multi-agent evaluation
• configs/agents/ag2-gpt4.yaml — Config for GPT-4 leaderboard runs
• configs/agents/ag2_agents.yaml — Aggregated config for agent_test.py
• requirements-ag2.txt — Isolated dependencies for AG2 environment

Modified files

• src/client/agents/__init__.py — Optional AG2 agent registration

Test plan

• 23 pytest tests pass (pytest src/client/agents/test_ag2_agent.py -v)
• Interactive REPL test via python -m src.client.agent_test --config configs/agents/ag2_agents.yaml --agent ag2-single
• AG2Agent returns correct responses from OpenAI API
• No impact on existing agents when AG2 is not installed

[faridun-ag2]

Hi @Xiao9905 @Longin-Yu!

Would love your review on this PR. It adds AG2 (formerly AutoGen) as a native agent client — the first multi-agent framework integration in AgentBench.

Key design decisions:

• Separate requirements-ag2.txt instead of modifying the main requirements.txt, since AG2 needs Python >=3.10 and pydantic v2 (incompatible with AgentBench's current deps)
• try/except ImportError in __init__.py so existing installations are unaffected
• 23 tests covering unit, integration, and e2e (with a fake task server)

Happy to adjust anything based on your feedback. Thanks!