persist Seq2Seq through keras load - save api

[stefanoterna]

Added a get_config override method to handle hidden_dim in LSTMDecoderCell and AttentionDecoderCell.

Moreover, the constructors of both classes relied on self.output_dim to set the value of hidden_dim if not supplied. But self.output_dim is not available until super constructor has been called. Fixed by accessing directly to kwargs, so that hidden_dim is be available when the super constructor calls the build_model method.