Fine-Tuning Vision-Language Models for Multimodal Polymer Property Prediction