Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models