Evaluating Structured Output Robustness of Small Language Models for Open Attribute-Value Extraction from Clinical Notes