Max_to cuda kernel bug with negative zero

[ViliamVadocz]

Code added in #370 contains a bug to do with negative zero.

Test case:

#[test]
fn test_max_negative_zero() {
    let dev: TestDevice = Default::default();
    let t = dev.tensor([[-0.0, 0.0], [0.0, -0.0], [-0.0, -1.0], [0.0, -1.0]]);
    let r = t.trace().max::<_, Axis<1>>();
    assert_eq!(r.array(), [0.0, 0.0, -0.0, 0.0]);
    let g = r.sum().backward();
    assert_eq!(
        g.get(&t).array(),
        [[1.0, 1.0], [1.0, 1.0], [1.0, 0.0], [1.0, 0.0]]
    );
}

This test passes on cpu, but with the test-cuda feature it gives this failure:

---- tensor_ops::max_to::tests::test_max_negative_zero stdout ----
thread 'tensor_ops::max_to::tests::test_max_negative_zero' panicked at 'assertion failed: `(left == right)`
  left: `[0.0, 0.0, -1.0, 0.0]`,
 right: `[0.0, 0.0, -0.0, 0.0]`', src\tensor_ops\max_to\mod.rs:144:9

This bug was mentioned in the StackOverflow page from which the code for the kernel was taken, along with a suggested fix: https://stackoverflow.com/a/72461459/10514840