In-context KV-Cache Eviction for LLMs via Attention-Gate