`torch.concat` is supported when reproducing results with docker

[pingzhili]

Many thanks for the kind sharing!!

When reproducing training results, the docker file specifies torch==1.8.1. However, torch at 1.8.1 had not introduced torch.concat yet, which will cause errors on this line of code:

https://github.com/EleutherAI/gpt-neox/blob/71df4d5017f9f4919566a11454fe3a507ffdc632/megatron/model/transformer.py#L432

Simply changing torch.concat to torch.cat can make it compatible.