Investigating Cross-Domain Capabilities of LLMs in Two-StageRecommender Systems