Evaluating the Robustness of Multilingual Transformer Models to Adversarial Attacks in Automatic Content Scoring