CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives