Beyond RAG: Task-Aware KV Cache Compression for Comprehensive Knowledge Reasoning