ci: drop custom NumPy builds for Python 3.15

[tirthpatel90]

Description

Fixes #2401

This PR removes custom NumPy build and wheel installation steps for Python 3.15 across CI workflow files (build-wheel.yml, test-wheel-linux.yml, and test-wheel-windows.yml) now that pre-built NumPy 3.15 nightly wheels are available.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mdboom]

/ok to test b33bb7c

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2411/
https://nvidia.github.io/cuda-python/pr-preview/pr-2411/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2411/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2411/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

It's going to be hard to evaluate this PR until the Python 3.15 is green on main again. Our self-built Numpy is segfaulting on import on main, and this upstream build of Numpy also is. It's great that the behavior matches, but we can't really evaluate that Numpy is working in CI until Numpy and/or CPython fixes the bug that is causing the segfault.

So we might just need to sit on this one for a while.

[seberg]

Thanks, but as the issue says, this should use the scientific-python-nightly-wheels not build from scratch.
Effectively, this just moved the build from build-time to test time. (Which pointed to it not being NumPy but rather an ABI change in CPython, but that is beside the point.)

Updating this will indeed not fix things since NumPy is also build with b2 but the main blocker for merging is that we should use the nightlies and not build from scratch.

[tirthpatel90]

Hi @seberg, I have updated the workflows to pass PIP_EXTRA_INDEX_URL for Python 3.15 so pip fetches pre-built nightly wheels from the scientific-python-nightly-wheels index instead of building from source. Thanks for pointing that out!

[seberg]

Unfortunately, I don't think this'll work, for two reasons:

1. We actually have narrow version pins numpy<2.5
2. I believe pip will still prefer the 2.5.x sdist and compile from scratch.

So, I think we need to relax the pin to <=2.6.0.dev0 (all numpy pre-releases). We could do that in env-vars via env vars by setting up what you have here plus PIP_ONLY_BINARY=numpy (to avoid pip prefering to build 2.5.x).
(We don't need PIP_PRE=1 if we change the pin, pip should eventually find that version if no other version has a wheel that works.)

Sorry for the creep, maybe I'll push it or follow-up too: The windows Python 3.15 jobs were currently deleted, can you try re-adding them now that we don't have to build NumPy anymore?
(if not that is fine, I'll follow up).