KV Shifting Attention Enhances Language Modeling