Use CuDNN GRU for 6x faster RNN computation

[bzamecnik]

Keras 2.1.0 offers keras.layers.CuDNNGRU, a drop-in replacement for keras.layers.GRU, that's backed by an optimized implementation from CuDNN.

Speedup for the RNN computation observed in nvprof is 6x, similar to what's reported elsewhere about CuDNNGRU. Note that it needs TensorFlow, CuDNN and an NVIDIA GPU. For CPU usage it has to be replaced by ordinary GRU layer. Saved weights are compatible. In practice it works well.

[tobegit3hub]

Thanks @bzamecnik and +1 for CudnnGRU.

[bzamecnik]

@tobegit3hub: We've already enabled the new CuDNN-compatible GRU to our internal OCkRE code, so I hope to port it to the public one soon.

[bzamecnik]

It looks that the upgrade to Keras 2 have not been published yet... This GRU needs Keras 2.1.5+, but 1.2.2 is in README.