Law of the Weakest Link: Cross Capabilities of Large Language Models