fix(release): validate packaged CUDA bundle

[xiaguan]

Summary

• validate the packaged binary without the CUDA action’s LD_LIBRARY_PATH, so $ORIGIN/../lib is exercised
• run the release build and packaging pipeline on release-related pull requests and manual dispatches
• split build, package, and tag-only publish into separate jobs
• upload the binary and packaged release candidate as short-lived Actions artifacts

Why

The tag-only single job required merging and moving the release tag for every packaging diagnostic. Package-only failures also repeated the full multi-architecture CUDA build. The split workflow makes pull requests the release dry-run and lets a failed package job be retried without rebuilding. Public release creation and attestation remain tag-only.

Evidence

• release run 32955220598 built successfully and showed ldd selecting /usr/local/cuda-13.0/lib64 because the CUDA action exports LD_LIBRARY_PATH
• local A/B: with that variable set, all three CUDA libraries resolve to the toolkit; with it removed, all three resolve through the packaged $ORIGIN/../lib RUNPATH
• the sanitized packaged binary reports pegainfer 0.1.1
• warm sccache run: 622 hits / 9 misses (99%); total job time 14m03s versus 21m19s cold, 7m16s faster (~34%)
• workflow YAML parsed with PyYAML
• bash -n scripts/package_qwen3_release.sh
• cargo fmt --check
• git diff --check
• pre-commit hooks passed

Signed-off-by: xiaguan 751080330@qq.com