Update `compareFloat16ULP` to align with DXC's implementation

[kmpeng]

ULP comparison support for Float16 was recently added based on the existing pattern in DXC's compareHalfULP function (see issue #81). There will likely be changes to the DXC implementation via PR #7393, which we should also update in our compareFloat16ULP implementation to maintain consistency with DXC.

Changes:

1. Handling +0 and -0 as equal
2. ULP differences are calculated with a monotonic ordering