LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings