Reduce duplicate CI package image downloads

[tpn]

Summary

Registry throttling has been failing installed-wheel/runtime checks before tests start. A PR revision currently pulls the same Python images in 12 separate jobs on both the PR event and its vetted push.

Run native wheel qualification once on the vetted push, alongside GPU checks. Ordinary PRs retain lint, CPU, and conda checks; ci-required still requires the complete native wheel matrix and GPU checks to succeed. Release builds retain full qualification.

Combine installed-wheel and runtime checks into six Python/architecture jobs. Each job pulls its existing pinned image once and runs the checks in separate fresh containers, reducing normal Python-image downloads from 24 to six per PR revision. Pulls have three bounded attempts; installation and test failures remain final. Python 3.12–3.14, both architectures, and existing Dragon coverage are preserved.

Using an authenticated shared cache remains a follow-up requiring CPU runner access. This change reduces traffic and handles transient pull failures; persistent registry quotas can still block CI.

Validation

• make ci-lint: passed.
• uv run --locked --extra dev pytest -q tests/test_release.py: 47 passed.
• Actionlint and extracted-shell ShellCheck: passed, excluding actionlint's unsupported existing concurrency.queue key.
• Executed the aggregate shell across 134 status cases; verified successful pulls, recovery, exhausted retries, and unretried test failures with stubbed commands.
• GitHub-hosted execution of the revised workflow remains pending.

Repository hygiene

• No credentials, private paths, local datasets, notebook outputs, or generated run artifacts were added.
• No large binary artifacts were added.
• Security-sensitive information is not included.
• Contributor CI behavior is documented.
• No new dependencies were added.

[coderabbitai]

Important

Review skipped

Auto reviews are limited based on label configuration.

🏷️ Required labels (at least one): 1

• ai-review

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cuPhoton/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 17b2d23a-cd45-46fd-b31c-562ca7334d61

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[melo-gonzo]

LGTM. This reduces duplicate image pulls while preserving package coverage and the required CI checks. Local release tests and CI status/retry checks passed. The two remaining CI failures are registry quota errors before tests start; a successful rerun is still needed.