Emit benchmark start/end markers for the tinybench plugin and the vitest
walltime runner, wrapping the measured function in a root frame.

The runner consumes the instrument-hooks FIFO stream in order and expects
SampleStart > BenchmarkStart > BenchmarkEnd > SampleEnd nesting per
benchmark, so the markers must land inside the sample window:

- vitest: emit the marker pair before stopBenchmark(), and move
  stopBenchmark() plus the markers into a finally block so a throwing
  benchmark cannot leave the profiler started-but-never-stopped
- tinybench: emit markers per task between start/stop instead of a single
  run-level pair; wrap the body in try/finally to keep start/stop balanced
  when a benchmark throws
- benchmark.js: bind wrapWithRootFrame/wrapWithRootFrameSync to the real
  implementations in the integ test's core mock

Add a regression test asserting both markers land between startBenchmark
and stopBenchmark in walltime mode.

Add benches/macos.bench.ts guarded with describe.skipIf(!darwin) so it
only runs on macOS, and wire the walltime-macos-test CI job to execute it
via a direct `node vitest.mjs bench --run macos` invocation.