Explore Better Relative Position Embeddings from Encoding Perspective for Transformer Models