Fix flaky TestChildProcessCleanup: poll for file growth, clean up proc in finally

The three TestChildProcessCleanup tests were flaky on Windows/macOS CI
because they used fixed anyio.sleep() durations (0.5s + 0.3s = 800ms)
to wait for nested Python interpreter chains to start writing to marker
files. On loaded CI runners, two Python startups + subprocess.Popen +
open() + first write() can exceed 800ms, causing assert 0 > 0.

Secondary bug: when the assertion failed, proc was never terminated
(finally only cleaned up tempfiles). The leaked subprocess was GC'd
during a later test, triggering PytestUnraisableExceptionWarning and
causing knock-on failures in unrelated tests.

Changes:

- Added _wait_for_file_size() helper: polls until getsize(path) exceeds
  a threshold, bounded by anyio.fail_after(10). Replaces all startup
  sleep+assert chains. Raises TimeoutError with a clear message instead
  of a confusing assert 0 > 0.

- Added proc cleanup to each finally block. Uses proc=None tracking
  after successful in-test termination to avoid redundant double-calls,
  and anyio.move_on_after() so cleanup timeout never masks the real
  failure.

- Removed the parent_marker check from test_basic_child_process_cleanup.
  NamedTemporaryFile(delete=False) already creates the file, so the
  os.path.exists() assertion was a no-op that never verified anything.
  Child writing proves parent started.

- Simplified shutdown verification: terminate_posix_process_tree already
  polls for process-group death, so the first post-termination sleep(0.5)
  was redundant. Reduced to a single 0.3s stability check (3x child
  write interval).

Result: 3 tests drop from ~7s to ~1.9s locally, 30/30 pass under
4-worker parallel load (flake-finder).

Closes #1775

Github-Issue: #1775

Author: maxisbey