fix(benchmarks): restore server call admission hot-path benchmark

[SunSi12138]

Fixes #756.

Summary

• Replace obsolete reflection over SharpLinkServer._connections with _connectionRegistry.SnapshotActive().
• Use the existing BenchmarkEnvironment.Server property instead of reflecting into its private _server field.
• Add --server-call-admission-smoke-test to execute GlobalSetup -> AcquireAndRelease -> GlobalCleanup, asserting balanced active calls; run it in allocation-gate CI.
• Keep the RPC session flush same-machine A/B base worktree compilable when it copies the PR-head Program.cs: copy the corresponding ServerCallAdmissionHotPathBenchmarks.cs as well (addresses review P2).
• Add .github/workflows/server-call-admission-benchmark-evidence.yml to execute the actual exact-head BenchmarkDotNet workload in Release, validate the generated ns/op and B/op results, and upload measurements and environment metadata (addresses review P3).
• Document the smoke command.

No changes to production src/, public API, or wire protocol.

Validation

• Branch diff verified against dev.
• PR-head smoke passed in the prior PR Fast run (per review).
• RPC Session Flush Performance Evidence rerun: https://github.com/SunSi12138/SharpLink/actions/runs/37880032450
• Exact-head full BDN evidence workflow: https://github.com/SunSi12138/SharpLink/actions/runs/37880032471
• Keep this PR in Draft until both workflows finish successfully and the BDN report shows measured latency and allocations.

Local commands

dotnet run --project test/SharpLink.Benchmarks/SharpLink.Benchmarks.csproj -c Release -- --server-call-admission-smoke-test
dotnet run --project test/SharpLink.Benchmarks/SharpLink.Benchmarks.csproj -c Release -- --filter '*ServerCallAdmissionHotPathBenchmarks*' --exporters json

[SunSi12138]

Review result: changes needed.

The benchmark setup fix itself looks correct and the new smoke passes on the PR head, but this PR currently breaks an existing same-machine A/B workflow.

RPC Session Flush Performance Evidence run 37878976018 fails deterministically while building the exact base benchmark harness: that workflow copies the PR-head Program.cs into the base worktree but does not copy the updated ServerCallAdmissionHotPathBenchmarks.cs, so the base build sees the new RunSmokeTestAsync() call without that member and fails with CS0117 before any measurements run.

Please update the A/B harness preparation so candidate-only CLI entry points do not make the historical/base worktree uncompilable, then rerun the workflow.

Separately, #756's completion definition asks for the actual BenchmarkDotNet run to produce ns/op and allocation data. The new smoke is useful for setup regression coverage, but it only executes Setup -> one AcquireAndRelease -> Cleanup. Before closing #756 / marking this ready, please attach at least one exact-head full ServerCallAdmissionHotPathBenchmarks run (or add an appropriate non-fast validation) showing that the complete BDN lifecycle and MemoryDiagnoser output work.