[master] CI/CD updates to be more stable

[josephevans]

This PR fixes a few issues with CI/CD:

1. When multiple processes are attempting to install a pip package at the same time, there is a race condition that causes them to fail intermittently. Since website s3 push and publish is not run inside a container, just use the awscli installed in the jenkins slave (which is up-to-date.)
2. Remove the onednn repository after installing onednn. This change prevents all CI pipelines from failing in case the onednn repository gets corrupt (or sync issues), since any apt calls will fail.
3. Update CUDA architectures built for Windows. Include 7.5 for Turing (which are on g4 instances,) so we can migrate to these instances for Windows CI.

[mxnet-bot]

Hey @josephevans , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [centos-cpu, sanity, windows-gpu, unix-cpu, windows-cpu, miscellaneous, website, unix-gpu, centos-gpu, clang, edge]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[josephevans]

I think the right approach would be to move the stuff into the container rather than adding another bandaid. The fact that anything runs on the host in the first place is already an issue.

Thanks for the suggestions. At this point, we do not have the resources to refactor this. If you would like to propose a solution and file a PR, we would greatly appreciate it.

[waytrue17]

LGTM, thanks