On the Encoder-Decoder Incompatibility in Variational Text Modeling and Beyond