eval: enforce one-tool-call response format on every turn

feat: add run_eval.py to Space (needed by eval_with_vllm.py for trained-model evals)