Support NumPy 1.23

[kmaehashi]

Tasks

• Read Release Notes.
• Fix for NumPy 1.23 #6807
• FlexCI
  • Update .pfnci/schema.yaml and .pfnci/matirx.yaml if needed. Support NumPy 1.23 #6820
  • Backport.
• Update Requirements (for NumPy/SciPy)
  • Bump the pinned minimum/maximum version in setup.py (see [info] NumPy/SciPy new version pinning recommendation #4773.) Support NumPy 1.23 #6820
  • Backport.
• Documentation
  • Update Requirements section in the Installation Guide (note: the baseline version must be updated only in the development branch). Also update docs/requirements.txt to use the latest versions. Support NumPy 1.23 #6820
  • Backport.

---

Test with NumPy 1.23rc2 failing:
https://ci.preferred.jp/cupy.linux.cuda-head/103003/
https://ci.preferred.jp/cupy.linux.cuda-head-stable/103006/

[kmaehashi]

https://numpy.org/devdocs/release/1.23.0-notes.html

Things that may be affected:

• Changing to dtype of a different size now requires contiguity of only the last axis
• keepdims parameter for average
• New parameter equal_nan added to np.unique
• 1D np.linalg.norm preserves float input types, even for scalar results

[kmaehashi]

Adding this one to rc milestone.

[kmaehashi]

As we are about to v11 branch cut, I think it's ok not to backport NumPy 1.23 things to v10. Even without the "official" support, users can safely use NumPy 1.23 + CuPy v10.

[kmaehashi]

As we are about to v11 branch cut, I think it's ok not to backport NumPy 1.23 things to v10. Even without the "official" support, users can safely use NumPy 1.23 + CuPy v10.

Discussed offline and agreed with this.