Dockerfile installation of torch and torchvision from s3, replacing original versions.

**What did you find confusing?**
In the [Dockerfile.gpu](https://github.com/aws/sagemaker-pytorch-training-toolkit/blob/master/docker/1.5.0/py3/Dockerfile.gpu), there is a point where torch and torchvision are uni-nstalled, to be replaced with the re-installed specialized version of both packages from:
- [here (torch)](https://pytorch-aws.s3-us-west-2.amazonaws.com/pytorch-1.5.0/py3/gpu/torch-1.5.0-cp36-cp36m-manylinux1_x86_64.whl) ,and
- [here torchvision](https://torchvision-build.s3.amazonaws.com/1.5.0/gpu/torchvision-0.6.0-cp36-cp36m-linux_x86_64.whl)

Around line 138, in the Dockerfile.

**Describe how documentation can be improved**
Why are these replaced? During image building, this process is time-consuming. What what happen if I removed these versions and kept the original?

Inspect docker/1.5.0/py3/Dockerfile.gpu around line 138 and review the linked torch and torchvision wheel sources. Document why the packages are replaced, what retaining the original versions would affect, and confirm the explanation against the image-building process.】【。

Written by the indexing model from the issue text.