Towards Robust Universal Information Extraction: Benchmark, Evaluation, and Solution