Alignment for Efficient Tool Calling of Large Language Models