Fix Matrix4x4.CreateReflection when D is not zero

[hez2010]

This is a correctness regression from .NET 8 which was introduced in #103527.
Fixed it and added a regression test for it.

Fixes #110050

[hez2010]

I think we should revert to the MathHelper.Equal as the sign of 0.0 in the result is not determined. I think as for this algorithm the sign of zero does not matter.

Plane plane = new Plane(0, 1, 0, 60);
Matrix4x4 actual = Matrix4x4.CreateReflection(plane);

Console.WriteLine(Equal(1.0f, actual.M11));
Console.WriteLine(Equal(0.0f, actual.M12));
Console.WriteLine(Equal(0.0f, actual.M13));
Console.WriteLine(Equal(0.0f, actual.M14));
Console.WriteLine(Equal(0.0f, actual.M21));
Console.WriteLine(Equal(-1.0f, actual.M22));
Console.WriteLine(Equal(0.0f, actual.M23));
Console.WriteLine(Equal(0.0f, actual.M24));
Console.WriteLine(Equal(0.0f, actual.M31));
Console.WriteLine(Equal(0.0f, actual.M32));
Console.WriteLine(Equal(1.0f, actual.M33));
Console.WriteLine(Equal(0.0f, actual.M34));
Console.WriteLine(Equal(-0.0f, actual.M41));
Console.WriteLine(Equal(-120.0f, actual.M42));
Console.WriteLine(Equal(-0.0f, actual.M43));
Console.WriteLine(Equal(1.0f, actual.M44));

(bool, string) Equal(float a, float b)
{
    if (a.ToString() == b.ToString()) return (true, "ok");
    return (false, $"expected: {a}, actual: {b}");
}

On .NET 9 with this PR:

(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(False, expected: -0, actual: 0)
(True, ok)
(False, expected: -0, actual: 0)
(True, ok)

On .NET 8:

(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)

On .NET 7:

(True, ok)
(False, expected: 0, actual: -0)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(False, expected: 0, actual: -0)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)

On .NET 6:

(True, ok)
(False, expected: 0, actual: -0)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(True, ok)
(False, expected: 0, actual: -0)
(False, expected: 0, actual: -0)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)
(True, ok)

Also, the precision of the existing test is an issue as well:

    System.Numerics.Tests.Matrix4x4Tests.Matrix4x4CreateReflectionTest01 [FAIL]
      Assert.Equal() Failure: Values differ
      Expected: -1.57142854
      Actual:   -1.57142889

I want to minimize the change given that this PR need to be backported to .NET 9. I'm reverting the changes to tests.

[tannergooding]

Determinism with -0 vs +0 is actually relevant and expected

The general issue is that the pre-existing CreateReflection01 test is “bad” and is essentially testing the output is the same as base math, which is faulty and not itself “correct”

I’d be fine with leaving the existing test using the helper to minimize PR churn, but the new test should be strictly doing the right thing here

[hez2010]

Test failures are unrelated.

@tannergooding PTAL

[tannergooding]

/backport to release/9.0-staging

[github-actions]

Started backporting to release/9.0-staging: https://github.com/dotnet/runtime/actions/runs/12018037447