Why Knowing Both Hops Is Not Enough: Understanding Two-Hop Generalization in Language Models