Fix spurious "thread death" TOCTOU race in python_worker

[NicoKiaru]

Summary

Fixes the sporadic spurious "thread death" task failures reported in apposed/appose#15.

Root cause

In Worker._process_input(), non-main tasks were wired up as:

task._thread = Thread(target=task._run, name=f"Appose-{uuid}")
task._thread.start()

task._thread is assigned before start() is called. The janitor thread _cleanup_threads() flags any task where task._thread is not None and not task._thread.is_alive(). A just-constructed, not-yet-started thread is not alive, so if the janitor's 0.05s poll lands in the window between Thread() construction and start(), it wrongly fails the task with "thread death" — even though the thread is about to run fine.

Fix

Build the thread into a local, start it, and assign task._thread only after start() returns, so the janitor can never observe a referenced thread that isn't yet alive:

t = Thread(target=task._run, name=f"Appose-{uuid}")
t.start()
task._thread = t

Reproduction & verification

Added test_thread_death_stress, which floods the worker with 16 threads × 200 tiny tasks (none of which can legitimately die), so any "thread death" is the bug.

• Before the fix: failed on every run (e.g. 5/3200, 8/3200, 4/3200 tasks failing, all with "thread death").
• After the fix: passes consistently (6/6 consecutive runs green).

Full tests/test_service.py Python suite passes with no regressions, including test_python_sys_exit (a genuine sys.exit in a task is now caught and reported as SystemExit: 123, which the test allows).