Lost in the Pipeline: How Well Do Large Language Models Handle Data Preparation?