[fix] Add the latest numpy version to requirements.txt

[anj-s]

What does this PR do?

Certain PRs (looks like forks) are running into mypy errors due to a bumped up numpy version that is being pulled in.

• Bumped up numpy to >=1.21 as a temporary fix. We ideally don't want to have any other dependency other than PyTorch.
• Since np.append is untyped to begin with I did not see any fix other than adding an ignore. Maybe I am missing something?
• I am still debugging how we are pulling in numpy 1.21 for PRs that are opened outside of the fairscale repo. PRs within the repo are using numpy 1.20. I tried invalidating the cache but that did not make a difference. PRs on different forks don't share a cache with the repo according to circleci. Maybe we want to consider that in the future.

Before submitting

• Did you have fun?
  • Make sure you had fun coding 🙃
• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
  • N/A
• Did you make sure to update the docs?
  • N/A
• Did you write any new necessary tests?
  • N/A
• Did you update the changelog? (if needed)
  • N/A

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

[min-xu-ai]

LGTM

[omry]

This is conflicting with fairseq on Python 3.6:
https://github.com/pytorch/fairseq/blob/master/setup.py#L196-L197

I am not sure what is the reason for the restriction there, but requiring 1.21 is breaking fairseq right now.

[omry]

More info:
numpy dropped Python 3.6 support in version 1.20.
As of now, fairseq still supports Python 3.6.

This diff makes Python 3.7 a hard requirement for fairscale, and thus for fairseq which depends on it.

Since the motivation here seems to be typing, I don't think it's worth changing the minimal supported Python version for fairseq and fairscale for it.

[min-xu-ai]

Thanks for the heads up, @omry! How about we add the requirement to requirements-test.txt? That should make our CI happy without require users to have newer numpy. @anj-s, what do you think?

[omry]

I get an uneasy feeling thinking that the CI will test on versions other than the official minimal versions.

[anj-s]

Thank you for the headsup @omry As you can see in the PR description I am still debugging why some users ran into this. This was only a quick fix to unblock users.
@min-xu-ai your suggestion sounds good. Opened a PR to make this change since we don't want to break fairseq.

[anj-s]

Merged PR. @omry Let me know if this unblocks you. Thanks!