Michelangelo: Long Context Evaluations Beyond Haystacks via Latent Structure Queries