Benchmark hidden sizes that give better efficiency and bigger batches

[ngimel]

Thank you for your RNN benchmarking work! Hidden size is somewhat arbitrary, so there is virtually no reason (other than being really tight on memory) to make it, say, 500 instead of 512. Underlying gemms in RNNs/LSTM's usually have better efficiency with power-of-two sizes. It would be good to have benchmarks for hidden sizes 128, 512, 1024. Also, pretty often people use bigger mini-batches than 20, so benchmarking batches of 32/64 also make sense.