Evaluating Zero-Shot Multilingual Aspect-Based Sentiment Analysis with Large Language Models