The Engineering Constraints of Distributed LLM Inference over the Open Internet