The RealHumanEval: Evaluating Large Language Models' Abilities to Support Programmers