TRUEBench: Can LLM Response Meet Real-world Constraints as Productivity Assistant?