Function launch_vector_add 

pub unsafe fn launch_vector_add(
    stream: &Arc<CudaStream>,
    func: &CudaFunction,
    a_ptr: u64,
    b_ptr: u64,
    c_ptr: u64,
    n: u32,
) -> Result<()>

Launch vector_add kernel: C[i] = A[i] + B[i].

This uses the safe cudarc launch_builder API with u64 device pointers. Since u64 implements DeviceRepr, no FFI conversion is needed.

Safety

All pointers must be valid CUDA device pointers to f32 arrays of length >= n.