onnxruntime

7c914a2a - [Python] Add an experimental CUDA 13 unified-wheel packaging pipeline (#33034)

[Python] Add an experimental CUDA 13 unified-wheel packaging pipeline (#33034) ### Description Adds a manually queued CUDA 13 packaging pipeline that produces one wheel per platform for Windows x64, Linux x64, and Linux ARM64. Each wheel contains the CPython 3.11–3.14 extension modules and shares a single copy of the native libraries; Linux also includes the 3.13t and 3.14t ABIs. The pipeline uses dedicated stage templates, tests the same Windows wheel on every supported Python version, and installs/imports the Linux wheel on every included ABI. Outputs are nightly-versioned pipeline artifacts. Existing packaging pipeline definitions remain unchanged, and the Linux build helpers retain their per-Python behavior unless explicitly invoked with `-m`. ### Motivation and Context Building a separate wheel for every Python version duplicates the large CUDA and TensorRT libraries. This experimental pipeline allows validation of unified wheels and native-library reuse without changing the existing packaging workflows. ### Validation - All 31 packaging tests pass, including regression coverage for consecutive ABI builds and unchanged per-Python wheel collection. The regression test reproduces stale auditwheel inputs before the fix; compilation and pip are stubbed while the shell build script and wheel merger run normally. - Nine focused shell checks pass for `-a`/`-m` forwarding, CUDA architecture overrides, ARM64 parallelism, and free-threaded dependency selection. - Repository lint passes for both Python files. - YAML template references and parameters, Bash syntax, embedded PowerShell syntax, and whitespace checks pass. - Full Azure packaging builds have not been run. --------- Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>