Testing the Limits of Fine-Tuning for Improving Visual Cognition in Vision Language Models