[dhiltgen]

This change exposes the set of CUDA compilation targets via device info so clients can determine if the build supports the discovered GPU's compute capability. This does not filter out unsupported devices to avoid ID reordering challenges if multiple devices are present, where only some are supported.

Clients can implement the compatibility logic outlined in https://docs.nvidia.com/cuda/cuda-programming-guide/01-introduction/cuda-platform.html#binary-compatibility and https://docs.nvidia.com/cuda/cuda-programming-guide/01-introduction/cuda-platform.html#ptx-compatibility to skip over unsupported devices.

Hypothetical example on a Blackwell GPU when compiled with MLX_CUDA_ARCHITECTURES=75-real;80-real;120a-real;120-virtual

{
    ...
    "architecture": "sm_120",
    "cuda_compilation_targets": "sm_75;sm_80;sm_120a;compute_120",
    ...
}

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: co-developed with coding agent

[zcbenz]

configure_file complicates build a lot, for simple case like this I prefer using target_compile_definitions to add a define directly.

[zcbenz]

This is no really a device info, it is independent of device. I think we can mirror the torch.cuda.get_arch_list API with mx.cuda.get_arch_list.

[dhiltgen]

Updated to use an explicit API instead of putting into device_info.