Evaluating Dialect Robustness of Language Models via Conversation Understanding