[zhangzhenghao]

Summary

• Add clap-cpp[cuda], clap-cpp[rocm] and clap-cpp[vulkan] as exact-version dependencies on disjoint native backend distributions, preserving the default CPU package and public Python API.
• Select the installed GPU backend automatically (or explicitly with CLAP_CPP_BACKEND), reject ambiguous/mismatched packages, and report missing GPU runtime dependencies rather than silently changing native modules.
• Enable GPU wheel CMake builds and connect CUDA/HIP to the existing cached native dense-matmul path; retain existing CPU precision/fallback behavior.
• Add backend-specific Python release jobs, sccache C/C++/CUDA/HIP launchers, auditwheel repair, source archives, GitHub release assets, and stable-release PyPI trusted publishing.

Verification

As requested, no tests were added or run.

• Built the CPU wheel through the default sdist-to-wheel path on Linux x86_64 / CPython 3.12.
• Built all four source distributions and inspected native sources, licenses, versions and extra dependencies.
• Inspected the CPU wheel's native module and shared-library dependencies (only standard platform runtime libraries).
• C++ syntax checks passed for CUDA, HIP and Vulkan compilation branches.
• Python 3.9 syntax, workflow actionlint and whitespace checks passed.

Actual GPU compilation/inference and the hosted release workflow have not been executed. GPU release wheels initially cover Linux x86_64; CPU wheels cover Linux x86_64/ARM64, macOS ARM64 and Windows x86_64, CPython 3.9–3.14. System GPU drivers/runtimes are required as documented.

First-release setup

Configure PyPI trusted publishers for clap-cpp, clap-cpp-cuda, clap-cpp-rocm and clap-cpp-vulkan, using gorse-io/clap.cpp, workflow python-release.yml, environment pypi. Stable releases publish to PyPI; prereleases attach GitHub assets only. No release or package has been published by this change.