Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformer Representations