A Study of Skews, Imbalances, and Pathological Conditions in LLM Inference Deployment on GPU Clusters detectable from DPU