Optimize export lookup and simple activation

[Andrew Arnott (AArnott)]

Summary

• cache exact runtime export lookups and initialized shared root values
• bypass general lifecycle state transitions for non-shared parts with no imports or OnImportsSatisfied callbacks
• avoid LINQ and empty-array allocations while resolving constructor arguments through the existing reflection lifecycle
• add BenchmarkDotNet coverage for shared, non-shared, constructor, property, complex, and ImportMany activation
• document the investigation, JIT tradeoffs, compiled activation experiment, and staged hybrid-plan ideas in PERFORMANCE.md

This branch does not add expression-compiled constructors, setters, or activation plans. That experiment is preserved separately on perf/compiled-activation-plans and has no pull request.

Performance

Three-run median from the IocPerformance scenarios (single-threaded, 500,000 iterations):

Scenario	MEF2	VS MEF	Ratio
Singleton	135 ms	207 ms	1.53x
Transient	128 ms	1,277 ms	9.98x
Combined	198 ms	5,300 ms	26.77x
Complex	499 ms	21,288 ms	42.66x
Property	1,121 ms	21,378 ms	19.07x
IEnumerable	791 ms	11,729 ms	14.83x

Compared with the original VS MEF measurements, this is approximately 30x faster for shared retrieval, 7x faster for simple non-shared activation, and 1.4-2.6x faster for the imported graph scenarios. The much larger imported-graph gains and their additional JIT cost are isolated on the topic branch.

Validation

• Microsoft.VisualStudio.Composition.Tests: 2,064 net8.0 and 2,107 net472 tests passed
• benchmark project builds without warnings for net10.0 and net472
• all six activation benchmarks execute under both target frameworks
• no new expression compilation remains in the pull request diff

Validation pipeline

vs-mef validation build 15110883