Are Smaller Open-Weight LLMs Closing the Gap to Proprietary Models for Biomedical Question Answering?