Text this: Zero-Shot Confidence Estimation for Small LLMs: When Supervised Baselines Aren't Worth Training