Backward ForgetMult + Bidirectional QRNN

[elmarhaussmann]

I had some time on the weekend and thought this might be fun. The PR implements a backward ForgetMult for CPU and CUDA and changes to QRNN and QRNNLayer to make them bidirectional. I tried to keep changes to the original code minimal without duplicating a lot of code.

I tested this with Pytorch 0.4.0 and Python 3.6.5. Gradient checks etc. pass. On preliminary results with IMDB movie reviews it looks like a bidirectional QRNN (2 layers, each with forward and backward, 256 hidden units) performs slightly better (~0.5% accuracy) than a unidirectional QRNN of same size (4 layers, 256 hidden units), but I haven't had enough time to finish experiments to be certain on this.

Let me know what you think and where the code still needs changes (I know it's not perfect in some places, especially QRNNLayer).

[salesforce-cla]

Thanks for the contribution! Before we can merge this, we need @elmarhaussmann to sign the Salesforce.com Contributor License Agreement.

[elmarhaussmann]

@Smerity: ping. No interest in merging? (I know it's old stuff now)

[dnbaker]

I'd be interested in having this merged in as well, though in the interim I'm glad to have the use of your fork. I imagine adding half/double support would also be worth doing.