eval: pass purpose=eval (server requires it with eval_sampling)

eval_adapter.py

@@ -50,6 +50,7 @@ def run_one(task_index, model):
         reward_key="correctness,reward",
         llm_url=VLLM_URL,
         model=model,
+        purpose="eval",   # required by the server when eval_sampling is passed
         eval_sampling={"temperature": 1.0, "top_p": 1.0},
         agent_step_limit=17,      # production eval cap from the FineEnvs Space
         agent_timeout_sec=600,