Use a block-reduction search in TensorPrimitives.IndexOfMin/Max/MinMagnitude/MaxMagnitude

[Mrnikbobjeff]

Summary

IndexOfMinMaxCore (shared by IndexOfMin, IndexOfMax, IndexOfMinMagnitude, IndexOfMaxMagnitude) tracks a running best value plus a vector of indices per lane. That costs a compare and two selects per element, needs the Size4Plus/Size2/Size1 variants so indices fit in lanes, and on AVX-512 hardware the blends of the running value go through mask-register conversions. IndexOfMin<int> over 65536 elements runs at about 5 GElem/s on a Zen 4 box, no faster than the .NET 10 package.

This PR replaces the loop with a two-pass block reduction, written against the operator interface so all four searches use it:

• Pass 1 reduces each block of 32 vectors to its best element with a pure vector loop (one load and one lane-wise Reduce per vector, two independent accumulators), then a horizontal Aggregate, and keeps the first block whose result beats the running result under the operator's strict Compare.
• Pass 2 scans only the winning block for the first element the result does not beat. That is the operator's own tie rule (first equal element; -0 before +0 for min; the sign preference for equal magnitudes), so no bitwise match is needed.

IIndexOfMinMaxOperator<T> gains Reduce (scalar and per vector width). Each operator forwards it to its aggregation operator (Vector.Min / Max / MinMagnitude / MaxMagnitude), which the previous code already used for its final horizontal step and bitwise match, so tie behavior is unchanged. Those reductions implement IEEE minimum/maximum and propagate NaN, so a block containing a NaN reduces to NaN and the first NaN of that block is returned; the per-vector NaN checks go away. Indices never live in vector lanes, so the per-element-size variants and ElementWiseSelect are removed (536 lines added, 597 removed).

Performance

BenchmarkDotNet 0.15.8, AMD Ryzen 7 7800X3D, .NET 10.0.5, IndexOfMin<int>, N = 65536 (256 KB), GElem/s:

Build	Increasing	Sparse hits	Random	Decreasing
main (this core, unmodified)	5.0	5.2	5.1	5.2
main, DOTNET_PreferredVectorBitWidth=256	8.0	7.9	9.1	8.7
This PR	36.7	36.7	36.6	35.6

The block loop compiles to two vpminsd zmm with memory operands per iteration; the result is independent of the input pattern and matches a hand-written Vector256 block-min argmin on the same machine (the L2 fill bandwidth for this working set). Inputs shorter than one vector use the scalar fallback.

Tests

• Adds multi-block length, NaN and signed-zero cases for all four searches (Helpers.TensorLengths stops at 256 elements, which is a single block at 256 bits).
• System.Numerics.Tensors.Tests filtered to *IndexOf* passes on the live net11.0 testhost, and the same test sources compiled against the net10.0 build pass on the 512-, 256- and 128-bit paths and with hardware intrinsics disabled (DOTNET_PreferredVectorBitWidth=256, DOTNET_EnableAVX2=0, DOTNET_EnableHWIntrinsic=0), for all 17 element types.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

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