rnn.py no longer breaks as expected when adding units

[charlesfrye]

In the RNN prediction video, a shape error is generated when the RNN outputs a 5-element hidden state when it should be outputting a 1-element hidden state.

Weirdly enough, in a more recent version of Keras, this error no longer occurs.

I suspect some weirdness is happening inside the loss function calculation.

Interestingly, the predictions on the validation data no longer line up, as can be seen in this project.