Disentangling Knowledge-based and Visual Reasoning by Question Decomposition in KB-VQA