MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly