Data Analysis in the Wild: Benchmarking Large Language Models Against Real-World Data Complexities