Optimizing KV Cache Eviction in LLMs: Adaptive Allocation for Enhanced Budget Utilization