판 이력 — HeadWiseKV: Budgeted Per-Head Cache Residency for Hybrid Long-Context Language Models | AIChainDay