CompareGreaterThan, CompareGreaterThanOrEqual, CompareNotGreaterThan, and CompareNotGreaterThanOrEqual have incorrect implementations

[tannergooding]

As per the title, the following methods have incorrect implementations and won't handle NaN inputs correctly:

• NI_SSE_CompareGreaterThan
• NI_SSE_CompareScalarGreaterThan
• NI_SSE_CompareGreaterThanOrEqual
• NI_SSE_CompareScalarGreaterThanOrEqual
• NI_SSE_CompareNotGreaterThan
• NI_SSE_CompareScalarNotGreaterThan
• NI_SSE_CompareNotGreaterThanOrEqual
• NI_SSE_CompareScalarNotGreaterThanOrEqual
• NI_SSE2_CompareGreaterThan
• NI_SSE2_CompareScalarGreaterThan
• NI_SSE2_CompareGreaterThanOrEqual
• NI_SSE2_CompareScalarGreaterThanOrEqual
• NI_SSE2_CompareNotGreaterThan
• NI_SSE2_CompareScalarNotGreaterThan
• NI_SSE2_CompareNotGreaterThanOrEqual
• NI_SSE2_CompareScalarNotGreaterThanOrEqual

For pre-AVX hardware, there isn't a hardware comparison mode that performs the given operations and so they are currently being implemented using the inverse operation.
However, using the inverse operation (such as converting GreaterThan to NotLessThanOrEqual) does not correctly handle the unordered relation (when one or both inputs is NaN).

These should instead be using a different inverse relation and swapping the operands. For example, CompareGreaterThan(a, b) should be emitted as CompareLessThan(b, a) as native does and as recommended by the x86 instruction manual: https://godbolt.org/z/ghwY5N

[tannergooding]

I found this while implementing #31193.

CC. @CarolEidt, @echesakovMSFT

[tannergooding]

This would be a technically breaking change to fix, but one we should likely take. Especially given the difference in behavior users will see for the 256-bit overloads being added in #31193 and people using the FloatComparisonMode enum overload.