pub fn numpy_to_tensor<B: Backend, const D: usize>(
    numpy_data: NpyData<'_, f32>,
    device: &B::Device
) -> Tensor<B, D>