Evaluating Large Language Models for Automated Evidence Synthesis in Neuroimaging AI: A Multi-Model Benchmark | Synapse