Fix aarch64 cpu/cuda include logic

[atalman]

Make sure with-cuda and with-cpu flags are correctly generating to aarch64 builds
Test:

python generate_binary_build_matrix.py --with-cpu disable  --with-cuda enable --with-rocm disable --operating-system linux-aarch64
{"include": [{"python_version": "3.9", "gpu_arch_type": "cuda-aarch64", "gpu_arch_version": "12.8-aarch64", "desired_cuda": "cu128", "container_image": "pytorch/manylinuxaarch64-builder:cuda12.8", "package_type": "wheel", "build_name": "wheel-py3_9-cuda-aarch6412_8-aarch64", "validation_runner": "linux.arm64.m7g.4xlarge", "installation": "pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128", "channel": "nightly", "upload_to_base_bucket": "no", "stable_version": "2.7.0", "use_split_build": false}]}

[vercel]

@atalman is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.