Accelerating recurrent neural network training using sequence bucketing and multi-GPU data parallelization