Add directory-style agent evaluation workflow and helper scripts

[wzzll123]

Motivation

• Provide a minimal, directory-based workflow for running code-generation agents against per-op tasks and evaluating outputs with the existing MultiKernelBench evaluator.
• Make it simple to prepare isolated task workspaces that contain only reference.py and agent_prompt.txt so agents write a single final_response.txt.
• Automate launching agents over many tasks and collecting structured evaluation summaries to speed up large-scale agent experiments.

Description

• Add scripts/agent_eval/prepare_workspaces.py which creates per-op task directories with reference.py, agent_prompt.txt, and a manifest.json, and supports --readonly to mark files read-only.
• Add scripts/agent_eval/launch_agent.py which runs an agent command template across task directories, supports placeholders ({prompt_file}, {prompt}, {workdir}), optional --auto-yolo behaviour, writes agent_run.json per task, and an overall agent_launch_summary.json.
• Add scripts/agent_eval/collect_results.py which invokes eval_single_runner.py on each agent-produced final_response.txt, writes eval_result.json per task, and emits aggregated JSON/CSV summaries.
• Add scripts/agent_eval/README.md and update top-level README.md with instructions and example invocations showing the three-step flow (prepare_workspaces.py, launch_agent.py, collect_results.py).

Testing

• No automated tests were added or run as part of this change.

---

Codex Task