CI: Extract shared CUDA Core wheel-build implementation after the CUDA 12 migration #3402

— codex GPT-6 on behalf of Ralf —

The current- and maintenance-toolkit Core wheel builds duplicate substantial steps in build-wheel.yml. Mike requested a tracker in #3318 so the copies do not drift.

After the CUDA 12 migration, extract a shared implementation and use it for both toolkit variants. Keep this as a focused, behavior-preserving follow-up.

Preserve toolkit switching, same-run bindings constraints, cleanup of stale outputs, ABI-specific wheel handling, and the final wheel-merging sequence. Validate both variants through the existing build/test CI.