Add langgraph based orchestrator

[dongwang218]

Why ?

Use langgraph to build state transition graph

How ?

also refactor to break p2p_agent into components.

Test plan

repeat test cases in #115

test the langgraph orchestrator

python -m matrix.agents.p2p_agents --config-name=coral_mmlu_pro.yaml max_concurrent_tasks=10 dataset.cut_off=10 resources.student_llm.matrix_service=gpt120b resources.teacher_llm.matrix_service=gpt120b resources.extractor_llm.matrix_service=8B output.path=$HOME/temp/coral_test_langgraph.jsonl resources.teacher_llm.sampling_params.max_tokens=10240 resources.student_llm.sampling_params.max_tokens=10240 orchestrator=coral_langgraph
[2026-03-04 02:53:09,800][__main__][INFO] - conv_err: 0.0000
[2026-03-04 02:53:09,801][__main__][INFO] - agreement: 1.0000
[2026-03-04 02:53:09,802][__main__][INFO] - agreement_correctness: 0.7000
[2026-03-04 02:53:09,803][__main__][INFO] - total_turns: 2.0000
[2026-03-04 02:53:09,803][__main__][INFO] - t_len: 303.2000
[2026-03-04 02:53:09,804][__main__][INFO] - s_len: 195.9000
[2026-03-04 02:53:09,804][__main__][INFO] - conv_len: 499.1000

python -m matrix.agents.p2p_agents --config-name=tau2_bench.yaml max_concurrent_tasks=10 dataset.cut_off=10 resources.user_simulator_llm.matrix_service=gpt120b resources.llm_agent_llm.matrix_service=gpt120b output.path=$HOME/temp/tau2_test_langgraph.jsonl domain=telecom orchestrator=tau2_langgraph
[2026-03-04 01:15:56,160][__main__][INFO] - rewards: 0.8000
[2026-03-04 01:15:56,160][__main__][INFO] - termination_user_stop: 1.0000
[2026-03-04 01:15:56,161][__main__][INFO] - termination_max_steps: 0.0000
[2026-03-04 01:15:56,161][__main__][INFO] - termination_too_many_errors: 0.0000
[2026-03-04 01:15:56,162][__main__][INFO] - termination_unknown_error: 0.0000

[yangli5t]

this could be deprecated?

[dongwang218]

the file still contains the entry point