Update the default RNNCell in tf 2.0 API.

1. Make the tf.nn.rnn_cell.RNNCell to be v1 only since it not properly behave
in v2, eg still use scope, etc.

2. Add the tf.keras.layers.AbstractRNNCell as the preferred
implementation of RNN cell. The cell has same method comparing to nn.rnn.RNNCell. The difference are base classes, and variable scope usage. This cell will be used in tf.contrib to replace the usage of LayerRNNCell as base class.

PiperOrigin-RevId: 235798594