[release/10.0] Non-zero result of Vector3.Cross with parallel vectors

[github-actions]

Backport of #119438 to release/10.0

/cc @tannergooding @pedrobsaila

Customer Impact

• Customer reported
• Found internally

Developers using Vector3.Cross(x, x) (that is the cross-product of parallel vectors) to be zero and would get a non-zero result.

Regression

• Yes
• No

This was regressed sometime in .NET 9 and wasn't caught due to it being non-LTS. The fix was introduced as part of an optimization to use MultiplyAddEstimate which allows the use of FusedMultiplyAdd on hardware that supports it, similarly to other vector math libraries (such as DirectX Math). This instruction is more accurate but can also cause a non-zero result in some edge cases due to the additional precision and inherent rounding error.

Testing

An explicit regression test was added.

Risk

Low risk. This simply changes out MultiplyAddEstimate(a, b, c) for an explicit (a * b) + c instead so that fusing can no longer occur.

[tannergooding]

CC. @jeffhandley for initial bar check

[jeffhandley]

This has my support for .NET 10 RC2. Functional fix for a regression introduced in .NET 9. If a customer reports being blocked by this in .NET 9 I would also support a reactive backport to .NET 9 servicing.

/cc @artl93

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-numerics
See info in area-owners.md if you want to be subscribed.

[artl93]

RC2. Regression.