fix(prefork): kill children that exceed per-job timeout (#81)

[pratyush618]

Summary

• Hung prefork children (infinite loop, blocking syscall, deadlock) used to wedge their reader thread forever and ignore @queue.task(timeout=…). The scheduler's stale-job reaper recorded the failure in the DB but the live process kept running, permanently losing that worker slot.
• Adds a watchdog thread that owns per-child Option<ActiveJob> slots. On deadline expiry it SIGKILLs + reaps the child, decrements in_flight, and emits JobResult::Failure { timed_out: true } with the same shape the reaper produces — so on_timeout middleware and JOB_TIMEOUT events fire identically to the thread-pool path. The slot acts as the single ownership token, so reader and watchdog can never double-complete the same job. The dispatcher's existing dead-child respawn loop brings the killed slot back on the next dispatch.
• No protocol changes — ParentMessage.timeout_ms and ChildMessage.Failure.timed_out already existed.

Closes #81.

Test plan

• cargo check/clippy --workspace clean for default, --features postgres, and --features redis
• cargo test --workspace — all 78 tests pass
• uv run python -m pytest tests/python/ — 464 passed, 9 skipped (incl. the pre-existing prefork basic-execution skip)
• uv run ruff check py_src/ tests/ and uv run mypy py_src/taskito/ clean
• New regression tests in tests/python/test_prefork.py:
  • test_prefork_kills_hung_task — while True: pass task with timeout=2 is killed within 12 s and fires on_timeout
  • test_prefork_no_timeout_unaffected — timeout=0 task runs to completion (watchdog must not kill it)
  • test_prefork_finishes_before_deadline — task completing before its deadline returns normally