Pseudo-Relevance Feedback Can Improve Zero-Shot LLM-Based Dense Retrieval