Add explicit numpy dependency

[garrison]

We are currently relying on numpy being installed as a transitive dependency (through, at the very least, Qiskit Terra). Even better would be to specific it explicitly, and to set the minimum version following #70 (comment).

TODO

• add a release note

[github-actions]

Pull Request Test Coverage Report for Build 5356659197

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 89.919%

---

Totals
Change from base Build 5348348321:	0.0%
Covered Lines:	2337
Relevant Lines:	2599

---

💛 - Coveralls

[caleb-johnson]

Isn't leaving it unpinned a little more robust since we don't rely on any unstable numpy features?

[garrison]

Isn't leaving it unpinned a little more robust since we don't rely on any unstable numpy features?

What do you mean by "more robust"?

I'll say a few things, though.

1. We should specify numpy as dependency since we import it directly. We should not rely on it being transitively depended on by Qiskit, even though there is no real risk of that changing.
2. Since we depend on it, it'd be nice to pin the minimum version to something sensible. Qiskit Terra depends on numpy>=1.17, so we could follow suit and do the same.
3. We know that the performance degrades on numpy versions earlier than 1.23 -- at least, if one uses CutQC to reconstruct the full probability distribution. I considered adding a warning, but it seems unlikely that everyone would notice it. I could think of no harm to bumping the minimum version of numpy to 1.23, so that's what I did here.

[caleb-johnson]

1. We should specify numpy as dependency since we import it directly. We should not rely on it being transitively depended on by Qiskit, even though there is no real risk of that changing.
2. We know that the performance degrades on numpy versions earlier than 1.23 -- at least, if one uses CutQC to reconstruct the full probability distribution. I considered adding a warning, but it seems unlikely that everyone would notice it. I could think of no harm to bumping the minimum version of numpy to 1.23, so that's what I did here.

Thanks, I like the justification of "we import it directly, so it is our dependency".

I agree, let's pin minimum to 1.23

[caleb-johnson]

LGTM 👍 thanks!