Layers with recurrent hidden state

Basically all layers which define get_rec_initial_extra_outputs.

Most (all) of them only when inside a rec loop, otherwise they behave just as if you would iterate over the time axis.

• WindowLayer: state [B,W,...]
• CumsumLayer: state which is the output itself
• RecLayer: like RnnCellLayer if unit is a str, or like subnet if unit is a subnet
• RnnCellLayer: state depending on unit option, e.g. LSTMStateTuple, or also often the output itself
• SelfAttentionLayer: k_left, v_left
• KenLmStateLayer: state, step, scores
• CumConcatLayer: state which is the output itself
• SubnetworkLayer: any from subnet
• BaseChoiceLayer, ChoiceLayer, DecideLayer, etc: choice_scores, choice_src_beams
• EditDistanceTableLayer: state, optional source_len
• MaskedComputationLayer: any from subnet, _output
• UnmaskLayer: t
• TwoDLSTMLayer: state, output, iteration