[libtorch] Remove cuda from the list of non-x86-windows triplets Default-Features

[JoergAtGithub]

Describe the bug
Only on non-x86 windows triplets, the default-features contain the GPU manufacturer specific CUDA support. On all other triplets the CPU only version is the default-feature and CUDA support is an optional feature for owners of an NVIDIA GPU.

Environment

• OS: "windows & !x86"

To Reproduce
Steps to reproduce the behavior:

1. ./vcpkg install libtorch
2. Default features fail on Windows x64/ARM systems without NVIDIA graphics drivers installed

Expected behavior

1. Make the CPU only version the default feature for all platforms. It's of cause slower in computition, but works everywhere without third-party drivers.
2. cuda should be a feature, that the user has to specify explicitly, if he want NVIDIA GPU support
3. Other GPU drivers like AMDs ROCm should be handled in the same way in future

Additional context
If we add other GPU vendor support (like AMDs ROCm) in the same way as default-feature, than I would need to install the drivers of all these GPU vendors in parallel on my system, to build the libtorch port. This concept does not scale.
There are libraries that depend on libtorch, it's difficult to propagate GPU vendor support, with this hardcoded triplet specific logic in the libtorch port.