Smaller but Better: Unifying Layout Generation with Smaller LLMs