CI: enable Windows ARM64 wheel builds for stable CUDA channels

Problem

The build-windows-arm64 job in .github/workflows/ci.yml currently runs only when:

needs.ci-vars.outputs.CUDA_BUILD_CHANNEL == 'prerelease'

This produced the Windows ARM64 cuda-bindings wheels for the CUDA 13.4 beta. When ci/versions.yml moves from the prerelease channel to stable for a full CUDA 13.4 release, the current condition will skip the Windows ARM64 build. The release workflow already downloads and publishes all matching component wheel artifacts, so a skipped build means there will be no stable win_arm64 wheel available to publish.

Proposed change

Enable the Windows ARM64 wheel job for supported stable CUDA versions. This may be done by removing the prerelease-channel check or replacing it with a capability/version guard that reflects Windows ARM64 toolkit availability.

Keep single-cuda-major: true until a prior CUDA-major Windows ARM64 toolkit and wheel set exists.

Also update the final CI status check so it expects the Windows ARM64 job to run for supported stable channels.

Acceptance criteria

• Tag-triggered CI for a supported stable CUDA release builds Windows ARM64 wheels for cuda-bindings and other supported compiled components.
• The stable Windows ARM64 wheels are uploaded as CI artifacts and are discoverable by ci/tools/download-wheels.
• A release dry run validates the Windows ARM64 wheels alongside the other platform wheels.
• Prerelease Windows ARM64 builds continue to work.
• Unsupported CUDA versions do not schedule a Windows ARM64 build.

Context

• Windows ARM64 build support was introduced for the CUDA 13.4 prerelease in #2437.
• The first published Windows ARM64 cuda-bindings wheels were released in 13.4.0b1.
• PyPI publication itself should not require special platform handling; the existing release workflow publishes every downloaded component wheel.