Rework the Cudnn LSTM implementation to get rid the extra input bias.

This heavily clear up the logic for save/load model since the weights for input bias is gone. The new implementation is mathematically same as the canonical LSTM.

PiperOrigin-RevId: 224531752