Layer-Specific Scaling of Positional Encodings for Superior Long-Context Modeling