The Bert Large training performance sometimes is wrongly calculated

[taotod]

The code below uses the final iteration training time to calculate the training performance.

https://github.com/IntelAI/models/blob/cdd842a33eb9d402ff18bfb79bd106ae132a8e99/quickstart/language_modeling/pytorch/bert_large/training/gpu/bf16_training_plain_format.sh#L57

If the final training iteration is at the end of the data file, it will be less than the expected batch size (16 or 32), then the final training iteration time will be very small (may be only half of the expected batch size, or less). Then this script will give the wrong performance data.

Suggest setting the parameter "drop_last" in the training code below to drop the final batch data of every data set file.
https://github.com/IntelAI/models/blob/cdd842a33eb9d402ff18bfb79bd106ae132a8e99/models/language_modeling/pytorch/bert_large/training/gpu/run_pretrain_mlperf.py#L904

[sramakintel]

@taotod could you submit a PR for us to review your suggestion?