Support packed sequence in our RNN implementation

[tobmi1]

Right now our RNN implementation (especially the dilated RNN) supports only tensor inputs. Ideally, it would also work with packed sequences like the vanilla torch RNNs