Rewrite TestChildProcessCleanup with socket-based deterministic liveness probe

Replaces the file-growth-polling approach with a fundamentally different
design: each subprocess in the tree connects a TCP socket back to a listener
owned by the test. Liveness and death are then detected via blocking I/O
operations that unblock on kernel-level events — zero sleeps, zero polling:

  1. await listener.accept() blocks until the subprocess connects
     (proves it started).
  2. After _terminate_process_tree(), await stream.receive(1) raises
     EndOfStream because the kernel closes all file descriptors of a
     terminated process. This is direct, OS-guaranteed proof that the
     child is dead.

Compared to the original file-watching approach:
- No tempfiles, no file I/O, no path escaping
- No fixed sleep() durations (the root cause of flakiness)
- No polling loops
- Failure modes are clear: accept times out = child never started;
  receive times out = child survived termination (actual bug)

Also fixes the process leak on assertion failure by adding proc cleanup
to each finally block (guarded by proc=None after successful in-test
termination to avoid noisy double-terminate).

Result: 3 tests drop from ~7.1s to ~0.5s, 30/30 pass under 4-worker
parallel load. Removes 80 net lines and the tempfile/path-escape deps.

Closes #1775

Github-Issue: #1775

Author: maxisbey