Self-GIVE: Associative Thinking from Limited Structured Knowledge for Enhanced Large Language Model Reasoning