Probing LLMs for Multilingual Discourse Generalization Through a Unified Label Set