ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases