Sometimes Painful but Certainly Promising: Feasibility and Trade-offs of Language Model Inference at the Edge