Use larger batch_size for more efficient computations

[bzamecnik]

Batch_size 32 is too small and not efficient. For a GPU with 8GB RAM, batch_size 256 is OK. A speedup observed going from 64 to 256 on this model was 2.55x. Note that even for smaller batch sizes CuDNNGRU (#2) is more efficient than ordinary GRU.