perf(tfork): replace Phase B polling with futex events- #9

[yiying-zhang]

Summary

• replace tfork restore-barrier 100 ms polling with FUTEX_WAIT on the exact
  observed nr_in_progress value;
• use the existing decrement-and-wake producer path;
• handle concurrent changes, signals, aborts, and spurious wakeups without a
  lost-wakeup window;
• preserve a monotonic 10-second failure deadline; and
• add an interleaved old/new Phase B benchmark harness.

Performance

At exactly 100 lightweight processes:

• isolated branch validation: median Phase B 822.833 ms to 306.195 ms
  (-62.8%), median command time 2,077 ms to 1,444 ms (-30.5%);
• final combined stack: median Phase B 770.573 ms to 266.309 ms (-65.4%),
  median command time 1,558 ms to 1,117 ms (-28.3%).

Validation

• every sample inherited the source sentinel;
• normal and abort wake paths remain bounded;
• the final combined stack passed all eight transactional fault boundaries,
  network-lock smoke, Gensee E2E, and the release soak;
• zero active bundles, runtime publications, tfork mounts, or D-state tasks
  remained.