Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs