RNN training now requires RNNCallback, but does not require RNNRegularizer; out and raw_out have moved to RNNRegularizer

[jph00]

Previously, it was necessary to add the RNNRegularizer callback to train an RNN, since that callback didn't only regularize, but also destructured the RNN's output as required for training. The latter step has now been moved to a new RNNCallback.

To make it easier to train RNNs, we've added an rnn_cbs which provides all callbacks needed for RNN training, optionally with regularization.