Revert DynamicRNNEstimator's ability to accept an instance of RNNCell.

As we move to make RNNCells subclasses of Layers, several things become clear:

1. Layers are stateful - once called, they keep track of their variables.
2. Estimators create a new graph for predict/fit/etc.
3. Items #1 and #2 are not compatible, because a Layer modified in .fit()
   cannot be used in .predict(): its variables and internal state tensors
   are associated with the graph created in fit() -- and cannot be used
   with the graph being created in predict().  Strange errors occur.

As a result, Estimators may never accept a Layer instance given this API
conflict.

This PR unblocks the move of RNNCell to be subclasses of tf.layers.Layer.
Change: 151735868