Add Windows ARM64 wheels

[ssss141414]

Summary

• build CPU-only Windows ARM64 wheels for CPython 3.12–3.14 on the native windows-11-arm runner
• recognize Visual Studio's ARM64 processor name and enable the MSVC AArch64 NEON path
• build the ARM64 package with Ruy and the compiler OpenMP runtime, without x64-only MKL, oneDNN, or CUDA dependencies
• test the installed ARM64 wheel, including multithreaded INT8 inference and repeated multithreaded float inference/context shutdown
• document the Windows ARM64 support range and CPU-only limitation

Dependency

This is currently stacked on google/ruy#377. That draft dependency disables Ruy's unsupported cpuinfo integration on Windows ARM64; without it, the final link fails with an unresolved cpuinfo_isa symbol. This PR should remain draft until the Ruy change is accepted and the submodule is repinned to its merged commit.

Ruy context shutdown

A multithreaded float Ruy workload exposed a Windows TLS teardown deadlock when a replica worker exited: all inner Ruy worker functions returned, but their std::thread::join() remained blocked while the outer worker was destroying its thread-local Ruy context. The context is now explicitly destroyed in ReplicaWorker::finalize() via destroy_context(Device::CPU), before the outer worker enters TLS cleanup. The ARM64 workflow recreates and destroys the context three times under a two-minute timeout to guard this path.

Validation

Validated on native Windows ARM64 with CPython 3.13.15, Visual Studio 2022 17.14, MSVC 19.44, and Windows SDK 10.0.26100.0:

• clean installed-wheel suite: 82 passed, 89 skipped
• multithreaded INT8 transliteration: passed
• three consecutive multithreaded float translator create/infer/destroy cycles: passed
• wheel tag: cp313-cp313-win_arm64
• _ext.cp313-win_arm64.pyd and ctranslate2.dll: PE machine AA64
• no non-system third-party DLL dependency; VCOMP140.DLL is supplied by the documented Visual C++ Redistributable