LLM Inference Beyond a Single Node: From Bottlenecks to Mitigations