Summary

When the parent pytest dispatcher's session watchdog fired
(--pto-session-timeout), the CI log only showed the timeout banner.
The actually-stuck child subprocess was killed silently, its captured
stdout discarded, and the child was left to be reaped as an orphan by
the runner cleanup — see run 25541961688
for the symptom (9.5 min of dead air, then [pytest] TIMEOUT, no
indication which case was stuck).

This PR makes the timeout path self-diagnosing.

Changes

• parallel_scheduler.py — emit [scheduler] START <label> pid=... devices=...
  at launch (not only at completion). The last START without a matching
  PASS/FAIL identifies the hung case at a glance.
• parallel_scheduler.py — expose _active_state while run_jobs
  is in flight so the parent's signal handler can reach the live job
  table.
• conftest.py — on session-timeout SIGALRM:
  1. send SIGUSR1 to every in-flight child (faulthandler, registered
     in pytest_configure, dumps all-thread Python + C stacks into the
     child's stdout — works even when the GIL is held by a native NPU
     call, which Python-level watchdogs cannot reach);
  2. let pumps drain ~2 s so the dumped stacks reach output_lines;
  3. print each in-flight job's tail buffer inside a ::group::HUNG ...
     block;
  4. call _terminate_all so children don't outlive us as orphans.

What the log will look like next time

Normal: every case has a [scheduler] START followed (out of order, in
parallel) by a ::group:: ... [PASS|FAIL ...] block.

On hang: the START for the stuck case has no matching PASS/FAIL, then
[pytest] TIMEOUT: ..., then a ::group::HUNG <label> pid=... elapsed=...
block containing the all-thread traceback the child wrote when it caught
SIGUSR1. The existing CI retry (any non-zero rc → re-run with the pinned
PTO-ISA commit) is unchanged: a PTO-ISA regression that surfaces as a
hang is exactly the case that retry path is meant to recover.

Notes

• pytest-timeout per-test guard is intentionally not enabled; the
  single-layer "parent catches everything via SIGUSR1" path covers all
  hang shapes (including native NPU deadlocks where pytest-timeout's
  watchdog thread can't fire). If hangs become frequent enough that
  burning the full session budget per incident is painful, adding
  timeout = N / timeout_method = "thread" to pyproject.toml is a
  pure increment on top of this PR.
• macOS has SIGUSR1; Windows does not. All new SIGUSR1 paths are
  guarded by hasattr(signal, "SIGUSR1").

Testing

• Local: trigger a hang test (time.sleep(9999)) under
  --pto-session-timeout 30; confirm the CI log shows
  ::group::HUNG ... with traceback, rc=124, no orphan python
  processes after exit.
• CI: PR validation against simulation runners (no actual hang,
  should look identical to before plus the new START lines).