OmniJARVIS: Unified Vision-Language-Action Tokenization Enables Open-World Instruction Following Agents | Read Paper on Bytez