Multimodal Large Language Models in Ophthalmology: Diagnostic Accuracy and the Risk of Metadata-Induced Confirmation Bias | Synapse