CI: matrix publish across macOS + Linux (x86_64/arm64)

[sqr00t]

Summary

Turn the CodeArtifact publish workflow into a build matrix so each supported platform produces its own wheel, while keeping the macOS publish as the gating leg.

Motivation

ahl_dwc is a compiled package (scikit-build-core + pybind11), so wheels are platform-specific. A single-runner workflow only ever produced one platform's wheel. A matrix lets us build for macOS and both Linux architectures from one release.

Scope

Leg	Runner	Toolchain	Behaviour
macOS arm64	macos-latest	Apple Clang + libc++	blocking; builds wheel + sdist, publishes
Linux x86_64	ubuntu-latest	GCC + libstdc++	non-blocking until the C++ fix
Linux arm64	ubuntu-24.04-arm	GCC + libstdc++	non-blocking until the C++ fix

Acceptance criteria

• Matrix over macOS arm64, Linux x86_64, Linux arm64 with fail-fast: false.
• macOS leg blocking; builds sdist (published once) + wheel; Linux legs build wheel-only.
• Linux legs continue-on-error so they don't block the macOS publish (they can't compile until C++ extension fails to compile under GCC/libstdc++ (blocks Linux wheels) #7).
• Publish endpoint kept in GitHub secrets (repo is public) — not committed to pyproject.toml.
• macOS build + publish confirmed green in CI (wheel uploaded; only a 409 on the pre-existing 0.1.0 sdist).
• Make publish idempotent so re-runs / already-published assets don't hard-fail (e.g. uv publish --check-url).

Dependencies / follow-ups

• C++ extension fails to compile under GCC/libstdc++ (blocks Linux wheels) #7 — fix the C++ so Linux compiles; then drop continue-on-error and switch Linux legs to cibuildwheel/manylinux (keeping scikit-build-core as the backend).