fix: tag tritonfrontend wheel with arch-specific platform tag (TRI-983)

[mc-nv]

What does the PR do?

Fix tritonfrontend Python wheel so it ships with a correct architecture-specific
platform tag instead of py3-none-any.

The root cause was a bdist_wheel.get_tag() override in src/python/setup.py that
hard-coded py3-none-any when --plat-name was not passed (which it never was).

Changes:

• src/python/setup.py — replace the bdist_wheel + get_tag() override with
  BinaryDistribution.has_ext_modules()=True, mirroring the fix applied to
  tritonserver in triton-inference-server/core. Setuptools now auto-derives the
  correct cpXY-cpXY-linux_<arch> tag from the current interpreter.
• src/python/build_wheel.py — add _repair_wheel_with_auditwheel() (ported from
  core/python/build_wheel.py) to upgrade the wheel to manylinux_2_X_<arch> for
  PyPI compatibility; add _detect_cuda_version() and _compose_version() for the
  +nv{release}.cu{cudaXY} local-version segment; add PEP 427 build-tag sourcing
  (CI_PIPELINE_ID → NVIDIA_BUILD_ID → BUILD_NUMBER) passed as
  --build-number=<N> to setup.py bdist_wheel; replace deprecated
  distutils.dir_util.copy_tree with shutil.copytree.

The container image continues to receive the linux_<arch> wheel from
generic/wheel/dist/; the manylinux wheel is written to generic/ for future PyPI
publishing.

^{CI (internal): [#49841381](http://tritonserver.local/ci/pipelines/49841381)
Resolves: TRI-983
Related: triton-inference-server/core#495}

Checklist

• PR title reflects the change and is of format <commit_type>: <Title>
• Changes are described in the pull request.
• Related issues are referenced.
• Populated github labels field
• Added test plan and verified test passes.
• Verified that the PR passes existing CI.
• Verified copyright is correct on all changed files.
• Added succinct git squash message before merging ref.
• All template sections are filled out.

Commit Type:

• fix

Related PRs:

triton-inference-server/core#495

Where should the reviewer start?

src/python/setup.py — BinaryDistribution class replaces the old get_tag() override;
src/python/build_wheel.py — _repair_wheel_with_auditwheel(), _compose_version(),
and the build-tag sourcing block.

Test plan:

Build inside the SDK container (requires auditwheel — added to Dockerfile.sdk in #8743):

python3 src/python/build_wheel.py \
  --dest-dir /tmp/whl \
  --binding-path <path-to-tritonfrontend_bindings.so>
ls /tmp/whl/              # manylinux wheel
ls /tmp/whl/wheel/dist/   # linux_<arch> wheel (for container install)
python3 -m wheel tags --list /tmp/whl/wheel/dist/*.whl
# Must NOT be py3-none-any

• CI Pipeline ID: 49841381

Caveats:

PyPI publishing CI jobs (py3-tritonfrontend-publish) are a follow-up task.

Background

Both tritonserver and tritonfrontend wheels were shipping py3-none-any despite
containing arch-specific .so extensions — causing Poetry/pip-tools lock file issues
for users on multi-arch setups. The tritonserver fix lives in triton-inference-server/core#495.

Related Issues:

• Resolves: TRI-983
• NVBug: 6098081 / JIRA: DLIS-8648