Encoder-Decoder or Decoder-Only? Revisiting Encoder-Decoder Large Language Model