From Reasoning to Generalization: Knowledge-Augmented LLMs for ARC Benchmark