clamp_min does not handle -inf correctly on Autodiff<NdArray> backend

[Karuso33]

The following code

    let x: Tensor<Autodiff<NdArray>, 1> = Tensor::from_floats([f32::NEG_INFINITY], &Default::default())
        .clamp_min(-1e8);
    assert!(x.into_scalar() >= -1e8);

currently panics (even though it clearly should not). Note that this does work correctly when Autodiff<NdArray> is replaced by either Autodiff<Wgpu> or by just NdArray.

I would expect clamp and clamp_max to have similar issues, but I have not tested it.