Add benchmarks for LSTMs

[hobofan]

LSTMs are very computationally heavy and produce state-of-the-art accuracy for some problems so adding models for them to benchmark would be interesting.