HateCOT: An Explanation-Enhanced Dataset for Generalizable Offensive Speech Detection via Large Language Models