perf: benchmark hot path and reduce logger dispatch overhead

[LimiNode]

Summary

• publish an immutable copy-on-write logger strategy snapshot to remove the per-call registry lock and temporary vector copy;
• make enabled and single_mode atomic and add a concurrent hot-path regression test;
• make async benchmark completion comparable: spdlog now waits for a worker-side flush marker, and the workload rejects incomplete sink delivery after adapter.flush();
• use one explicit blocking queue capacity for both adapters and record it in CSV output;
• reject zero queue capacity and refuse to append to an older latency CSV schema;
• align benchmark methodology with the prepared-message/direct-dispatch contract, pin spdlog 1.17.0, use Release CI benchmarks, and mark the old fixture as legacy.

Validation

• MSVC Release build of the C++17 benchmark targets, with the library core exercised by the C++11 matrix jobs;
• logit_bench_flush_test passed with an async queue capacity of 8;
• async null and file smoke runs passed for LogIt++ and spdlog;
• queue-capacity and CSV-schema validation paths were checked locally;
• fresh GitHub CI is green across Linux/Windows/macOS, C++11/17, ASAN/UBSAN, TSAN, MDBX, vcpkg, Emscripten, benchmark-asan, and the executed spdlog flush regression.

The PR intentionally does not claim a public LOGIT_INFO(...) benchmark or lock-elision for arbitrary custom backends. The legacy 2025 fixture is not presented as a current numeric comparison because its dependency/toolchain and async drain provenance were not recorded.