Unused linear layers in pointer-generators

[kylebgorman]

The GRU pointer-generator implementation includes a linear layer called merge_h, which presumably is intended to be used to "merge" hidden state from the source and features encoders; the LSTM one also includes an additional layer called merge_c, presumably for the cell state. However, these are not actually ever used, and as far as I can tell, they never were (I looked back a bit) and should either be removed or perhaps wired in.

@Adamits, I have assigned this to you because I'm frankly afraid to touch this module much, but maybe you'd be comfortable addressing this.