tighten wheel size limits, expand CI-skipping logic, other small build changes

[jameslamb]

Description

Proposes a batch of miscellaneous build / packaging / CI changes.

Changes

Tightens wheel size limits

Contributes to rapidsai/build-planning#219

CUDA 13 wheels can be as much as 50% smaller than the equivalent CUDA 12 wheels, because of some new compression features in nvcc.

To ensure surprising package-size growth is caught in CI, this PR tightens the limits in the following ways:

• setting different limits based on CUDA major version where package size differs significantly
• setting libcuvs limits to {compressed_size} + 40Mi, rounded to the nearest 5Mi
• setting cuvs limits to {compressed_size} + 10Mi, rounded to the nearest 5Mi

Expands CI-skipping logic

Contributes to rapidsai/build-planning#243

Tries to avoid unnecessary CI runs by making the CI-skipping rules finer-grained. For example, PRs that only touch .pre-commit-config.yaml should now not require any runners with GPUs 😁

This is extra important in this project which builds so many language bindings and where some test jobs take 2+ hours to run. As of this PR, no more waiting for all the C++ test jobs to run when you only change Go or Java code 😁

Removes reliance on pypi.nvidia.com

Contributes to rapidsai/build-planning#241

git grep -i -E 'pypi\.nvidia\.com'
git grep -i -E 'rapids\-init\-pip'

And removed/updated all relevant references. This project does not need any wheels from pypi.nvidia.com at build-time or runtime, it can safely avoid searching that index.

Its own packages are not available on pypi.org yet, so references in documentation are preserved.

Enforces PEP 639 license metadata in pyproject.toml

Contributes to rapidsai/pre-commit-hooks#95

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[jameslamb]

/ok to test

[bdice]

/merge

[bdice]

/merge