Defective Task Descriptions in LLM-Based Code Generation: Detection and Analysis