Distilling Instruction-following Abilities of Large Language Models with Task-aware Curriculum Planning