Φ-Bench: Can Large Language Models Engineer the Infrastructure That Powers Them?