Reduce subscriber contention with bounded spinning

[prezaei]

When subscribers contend for the read lock, blocking Dequeue currently yields and progresses to semaphore waits of up to 10 ms. Use bounded SpinWait retries before a 5 ms semaphore wait, resetting the spin budget after waiting. TryDequeue still makes one read-lock attempt, and the existing publisher/subscriber crash recovery remains intact. Also observe cancellation while waiting for a publisher to finish writing a reserved message.

Add eight regression cases covering contended reads, expired read locks, cancellation, and exactly-once delivery with concurrent publishers and subscribers. Add a reproducible concurrent-subscriber benchmark, support benchmark selection from the command line, and update existing benchmark jobs to .NET 10. This branch includes current main and the queue lifetime fix.

Validation:

• macOS and Linux full suites: 101 passed, 3 platform-specific skips each.
• Release solution build: zero warnings or errors.
• CI passed on Windows, macOS, and Linux for commit 03ef4d8: https://github.com/cloudtoid/interprocess/actions/runs/34725696901

The benchmark refresh corrects the enqueue operation count and ensures the ring-wrap workload actually crosses the buffer boundary. All seven benchmarks were run directly on an Apple M5 Max with .NET 10, three warm-up iterations, and eight measured iterations. The README includes the native Mac results and full reports; its historical Windows and Linux measurements are unchanged.

Native Mac workload	Mean per operation
Enqueue, 3 bytes	182.3 ns
Enqueue + dequeue, 3 bytes, reused buffer	210.0 ns
Enqueue + dequeue, 3 bytes, new result array	214.9 ns
Enqueue + dequeue, 50 bytes, reused buffer	214.6 ns
Enqueue + dequeue, 50 bytes, ring-wrap workload	223.8 ns
Concurrent delivery, 8 bytes, 1 subscriber	246.5 ns
Concurrent delivery, 8 bytes, 4 subscribers	344.7 ns

These are in-process microbenchmarks. Concurrent delivery measures amortized time per message, including worker startup and completion, rather than individual message latency. Earlier native Mac comparisons against main measured 253.3 → 241.0 ns with one subscriber and 1,505.4 → 342.0 ns with four; those used only three measured iterations and the contended baseline was noisy.

Reproduce all cases from the repository root:

dotnet run --project src/Interprocess.Benchmark -c Release -- --filter '*' --warmupCount 3 --iterationCount 8 --artifacts BenchmarkDotNet.Artifacts