Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models