Engagement Undermines Safety: How Stereotypes and Toxicity Shape Humor in Language Models