HumanVBench: Probing Human-Centric Video Understanding in MLLMs with Automatically Synthesized Benchmarks