perf: net10.0 zero-alloc hot paths in NamespaceTree + BenchmarkDotNet project

[Rolling2405]

Summary

This PR adds tiered, backward-compatible performance optimizations to NamespaceTree — the core hot path called for every scanned type during dependency search — taking full advantage of .NET 8/9/10 APIs while keeping
etstandard2.0 entirely unchanged.

Approach: #if NET*_OR_GREATER tiers, no new TFMs

No new TFMs are added. The library already targets
etstandard2.0;net10.0. All new code is gated with #if NET8_0_OR_GREATER, #if NET9_0_OR_GREATER, or #if NET10_0_OR_GREATER so:

• netstandard2.0 consumers: zero changes, zero risk
• net10.0 consumers: all tiers active automatically
• net8.0/net9.0 consumers: respective tiers active (since NET8_0_OR_GREATER is defined for all of 8, 9, and 10)

Optimizations by tier

Unconditional (all TFMs)

• NormalizeString fast-path: str.IsNormalized(FormC) is allocation-free; Normalize() always allocates. .NET type names are virtually always ASCII → NFC, so this is a near-100% zero-allocation fast-path.

#if NET8_0_OR_GREATER

• FrozenDictionary build→freeze pattern: Node builds child nodes into a Dictionary<string, Node> during NamespaceTree construction, then Freeze() is called at end of constructor, converting every node's dictionary to a FrozenDictionary. Read path gets SIMD-accelerated lookups; write path (construction only) stays a regular Dictionary. StringComparer.Ordinal used explicitly to enable GetAlternateLookup later.
• SearchValues: Created once as static readonly, replaces char[] in IndexOfAny — SIMD-accelerated character scan on modern runtimes.

#if NET9_0_OR_GREATER

• CollectionsMarshal.GetValueRefOrAddDefault on GetOrAddNode write path: single hash probe for insert-or-get instead of TryGetValue + Add (2 probes).
• FrozenDictionary.GetAlternateLookup<ReadOnlySpan>() on TryGetNode read path: zero-allocation span lookup against the frozen dictionary — avoids Substring allocation AND the NormalizeString check on the hot read path.
• Inline-span GetAllMatchingNames for both the string and TypeReference overloads: ullName.AsSpan(start, len) passed inline as a method argument (not stored as a local) to TryGetNode(ReadOnlySpan), which is CS4007-safe in iterator methods. Eliminates Substring allocations on the critical read hot path.

Files changed

• sources/NetArchTest/NetArchTest.csproj — LangVersion made conditional: latest for net10.0 (needed for C# 13 �llows ref struct in GetAlternateLookup), 11 for other TFMs.
• sources/NetArchTest/Dependencies/DataStructures/NamespaceTree.cs — tiered optimizations as described above.
• ests/NetArchTest.Benchmarks/ — new BenchmarkDotNet project targeting
  et8.0;net10.0. [MemoryDiagnoser] included. Two public-API benchmarks over the dependency search hot path. [SimpleJob(RuntimeMoniker.Net80)] + [SimpleJob(RuntimeMoniker.Net100)] so the difference is directly measurable.

Test results

353 passed, 0 failed, 2 skipped on both
et8.0 and
et10.0.

Backward compatibility

etstandard2.0 compilation: zero changes (all new code is inside #if NET*_OR_GREATER blocks).

• Public API surface: unchanged. All optimizations are internal to NamespaceTree and Node.
• IsAotCompatible / IsTrimmable: not set (deliberate, see PR feat(aot): AOT/trim annotations on dependency-search surface + NativeAOT smoke #12 for rationale).