Add CI for Windows CUDA build

[zcbenz]

This PR does a rewrite of Linux build's GitHub Actions workflow to support Windows.

Simplified overview for daily testing:

  build_and_test:
    strategy:
      matrix:
        os: ['Linux', 'Windows']
        arch: ['x86_64', 'aarch64']
        toolkit: ['cpu', 'cuda-12.6', 'cuda-12.9', 'cuda-13.0']
    steps:
      - uses: actions/checkout@v7
      - uses: ./.github/actions/setup
      - uses: ./.github/actions/build
      - uses: ./.github/actions/test-linux
      - uses: ./.github/actions/test-windows

---

Simplified overview for release build:

  build_frontend:
    strategy:
      matrix:
        os: ['Linux', 'Windows']
        arch: ['x86_64', 'aarch64']
        python-version: ['3.10', '3.11', '3.12', '3.13', '3.14']
    steps:
      - uses: actions/checkout@v7
      - uses: ./.github/actions/setup
      - uses: ./.github/actions/build-wheel
        with:
          build-backend: false
      - uses: ./.github/actions/test-linux
      - uses: ./.github/actions/test-windows
      - uses: actions/upload-artifact@v7

  build_backend:
    strategy:
      matrix:
        os: ['Linux', 'Windows']
        arch: ['x86_64', 'aarch64']
        toolkit: ['cpu', 'cuda-12.9', 'cuda-13.0']
    steps:
      - uses: actions/checkout@v7
      - uses: ./.github/actions/setup
      - uses: ./.github/actions/build-wheel
        with:
          build-frontend: false
      - uses: actions/upload-artifact@v7


  test_wheel:
    needs: [build_frontend, build_backend, build_mac_wheels]
    strategy:
      matrix:
        os: ['Linux', 'Windows', 'macOS']
    steps:
      - uses: ./.github/actions/test-wheel

---

For the release build the new test_wheel job will download the uploaded wheels into a fresh env and run MLX tests with them.

The nightly workflow is removed, instead the release workflow will run nightly to test the wheels.

This PR also includes some changes to fix compile cache misses.

Note that at the time of writing this PR uses the free Windows runners and we will switch to large runners later as fresh builds would take too much time on the free runner.