Mucko: Multi-Layer Cross-Modal Knowledge Reasoning for Fact-based Visual Question Answering