Star Attention: Efficient LLM Inference over Long Sequences