OccuBench: Evaluating AI Agents on Real-World Professional Tasks via Language World Models