Ensure FMA optimizations kick in under embedded broadcast

[tannergooding]

Follow up to #116804. The logic is correct, but it missed a case where embedded broadcast containment would interfere with the FMA optimization and cause it to be skipped entirely.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull Request Overview

This PR ensures that FMA (Fused Multiply-Add/Subtract) optimizations are correctly applied even when constant vectors are candidates for embedded broadcast. It refactors the intrinsic classification and refines the containment logic to account for negative-zero broadcasts interfering with FMA.

• Unified handling of AVX2/AVX512 FMA intrinsics in LowerFusedMultiplyOp
• Updated containment logic in ContainCheckHWIntrinsic to skip embedded broadcasts when better FMA transformations exist
• Added a new predicate OperIsVectorFusedMultiplyOp to identify FMA intrinsics

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/coreclr/jit/lowerxarch.cpp	Refactored FMA case switches, adjusted operand indexing, and enhanced embedded-broadcast checks
src/coreclr/jit/gentree.h	Declared OperIsVectorFusedMultiplyOp to flag FMA intrinsics
src/coreclr/jit/gentree.cpp	Defined GenTree::OperIsVectorFusedMultiplyOp() with documentation and intrinsic ID checks

Comments suppressed due to low confidence (2)

src/coreclr/jit/lowerxarch.cpp:9881

• New logic skips embedded-broadcast folding for FMA when negative-zero constants are involved. Add targeted unit tests to cover both the ordinary embedded-broadcast path and the skipped path to prevent regressions.

                            containedOperand->IsCnsVec() && node->isEmbeddedBroadcastCompatibleHWIntrinsic(comp);

src/coreclr/jit/lowerxarch.cpp:1509

• The operand index was changed from Op(1) to Op(2). Verify that this matches the intended third operand of the FMA node and doesn't introduce an off-by-one reference.

                GenTree* argOp = hwArg->Op(2);

[tannergooding]

CC. @dotnet/jit-contrib, small improvement that with some diffs in common math helpers ensuring we can do the valid FMA transformations when embedded broadcast exists

[kunalspathak]

LGTM