[wasm] TensorPrimitives.Tan returns +0.0 for range-reduced Half inputs when Vector128 is accelerated without FMA #135117

Description

TensorPrimitives.Tan for Half returns +0.0 for inputs that need range reduction on WebAssembly when Vector128 is hardware accelerated (seen in the CoreCLR WASI composite ReadyToRun lane; the WASI interpreter lane, where Vector128 isn't accelerated, passes).

This lines up with #134914, which rewrote Tan range reduction and added these tests. When neither Fma.IsSupported nor AdvSimd.Arm64.IsSupported, the new path does the non-fused multiply-adds and, if |f| < dn * TanNonFusedReductionThreshold, falls back to ApplyScalar<TanOperatorSingle>(x). WebAssembly has accelerated Vector128 but no FMA, so it's essentially the only configuration that takes this branch (x64 with FMA and Arm64 take the fused path; non-accelerated configurations take the scalar path). Half is widened to float and goes through TanOperator<float>. Only the Half instantiations fail; float/double pass.

cc @tannergooding

[FAIL] System.Numerics.Tensors.Tests.HalfGenericTensorPrimitives.Tan_RangeReduction(length: 2, inputKind: 0)
Expected: -407.25
Actual:    +0.0
[FAIL] System.Numerics.Tensors.Tests.TensorTests.TensorTanPreservesScalarAccuracy(layout: 2, columns: 33)
Expected: -407.25
Actual:    +0.0
[FAIL] System.Numerics.Tensors.Tests.HalfGenericTensorPrimitives.SpanDestinationFunctions_InPlace(... Tan ...)
Assert.All() Failure: 6 out of 257 items in the collection did not pass.
Expected: -0.92969
Actual:    +0.0

34 failures in System.Numerics.Tensors.Tests (Helix jobs e1cac54a-421c-439e-b0cd-8d98aff74c83 and eccd7d69-de48-4cc1-87fd-6036eb264d50, #134813). Tan_RangeReduction, SpanDestinationFunctions_InPlace and TensorTanPreservesScalarAccuracy are disabled via ActiveIssue on wasm with accelerated Vector128 (all element types) in #134813; the right scoping or fix is left to the area owner.