EASE: Practical and Efficient Safety Alignment for Small Language Models