#2063 Add Windows ARM64 RTX Spark support

说明

## Title
<!-- If Applicable, reference the GitHub issue -->
Add Windows ARM64 RTX Spark support

Issue: N/A

## Summary
<!-- Quick Summary of changes -->
Add experimental Windows and Linux ARM64 support, including Windows RTX Spark unified-memory detection and a native Windows ARM64 CUDA build path. ARM64 uses ggml-cpu for native/IQ experts; the x86-only legacy Q2_0 model remains unavailable.

## What changed
<!-- Specifics on files changed, and what changes were made there -->
- Detect ARM64 hosts and select ARM-safe CPU sources, spin waits, memory fences, and ggml-cpu expert fallbacks.
- Locate a CUDA-compatible Visual Studio installation with `vswhere`, requiring the ARM64 C++ tools without pinning one Visual Studio edition, and use native ARM64 LLVM/Clang plus MSVC as NVCC's host compiler.
- Account for RTX Spark's dedicated carve-out and Windows shared GPU memory when choosing models and reporting usable memory.
- Query the Windows DXGI local-memory budget for integrated NVIDIA devices.
- Add setup policy tests for ARM64 model selection, compiler discovery, and RTX Spark memory accounting.

## Extra Notes
<!-- Any extra notes, delete if there are none -->
Validation:
- `python tools/test_setup_arm64.py` - passed, 9 tests.
- `python tools/test_setup_choices.py` - passed, 47 tests.
- `python tools/test_setup_prompts.py` - passed, 7 tests.
- `python tools/test_setup_oldcpu.py` - passed, 7 tests with 1 skipped.
- `python tools/test_setup_golden.py` - passed, 4 tests.
- `python tools/test_setup_unsloth.py` - passed, 45 tests.
- `cmake --build build-arm64 --target strata -j 4` - passed with CUDA 13.4, native ARM64 clang-cl, and the MSVC ARM64 NVCC host compiler.

The x86 CUDA, HIP, and SYCL engines were not built on this Windows ARM64 machine. Existing x86 setup profiles remain covered by the setup tests above.