fix(local-llm): CUDA build that ships its runtime on Windows, and Windows on ARM engine (ATO-244, ATO-252)

[sosidudku1]

On Windows the agent picked llama-turboquant-windows-x64-cuda-13.3.zip for any NVIDIA driver with CUDA 13.x. In release turboquant-6df272c that zip came without the CUDA runtime (cudart64_13.dll, cublas64_13.dll, cublasLt64_13.dll). So on a machine without the CUDA Toolkit, ggml-cuda.dll could not load and llama-server silently ran on the CPU. On an RTX 3080 Ti with driver 591.44, prompt processing ran at about 16 tok/s and the first turn took about 7 minutes.

The engine packaging was fixed separately (AtomicBot-ai/atomic-llama-cpp-turboquant-nightly#4, release turboquant-ad5ad5f). This PR makes the agent safe even when a zip arrives incomplete.

What changes

• Selection. With driver CUDA ≥ 12.4 the agent picks cuda-12.4 (newer drivers run older runtimes). On Blackwell (compute capability ≥ 10, read with nvidia-smi --query-gpu=compute_cap) it picks Vulkan, because the 12.4 build has no code for it. Below 12.4, or with no NVIDIA driver, it picks Vulkan, as before. cuda-13.3 is no longer auto-picked but can still be pinned with localModels.managed.backendVariant.
• Completeness guard. Under auto, an installed CUDA build without cudart64_*.dll next to ggml-cuda.dll counts as stale and is replaced through the existing auto-update at managed start. Existing broken installs are fixed without manual steps.
• No re-download loop. If a CUDA zip itself arrives without its runtime, downloadBackend refuses it and installs the same release's Vulkan zip. It records refusedCudaAsset in backend-version.json and tries the CUDA zip again only on a newer release.
• Tests cover the selection table, the stale checks, and the refuse-and-fall-back path. README and the backendVariant comment are updated.

The CPU zip is still never picked by detection; cpu-backend-fallback.ts only learns to never fall back on win32-arm64.

Windows on ARM (ATO-252)

• resolvePlatformAsset no longer throws on win32-arm64. It resolves to llama-turboquant-windows-arm64-cpu.zip (engine release turboquant-9ca222f, CPU only); backendVariant and the NVIDIA detection are ignored there, and there is no CPU fallback loop.
• When no release carries that asset, downloadBackend and models update say plainly: "Local models are not available yet for Windows on ARM. Use a cloud model instead." models update exits 1 instead of reporting "backend unchanged".

Checks

• On the MacBook Pro: agent lint and build clean, the whole vitest suite 12 925 passed with 0 failed (with the ARM part).
• The same change runs in the desktop batch (Desktop 06.10 batch with Windows on ARM (ATO-252) #614): full desktop smoke and e2e (8/8) on macOS, plus the smoke suite on Windows x64 and Windows on ARM runners (738 passed on each, no unexpected failures).
• On the reporter's PC (RTX 3080 Ti) the test build loads the model on the GPU.