Rewinding One Request in a Continuously Batched LLM ServerBanning specific whole-phrases in a sequence