perf(session-memory): use native timed join for generated reader

[8Dionysus]

Summary

• replace the generated-reader polling loop's fixed 50 ms join slices with one native Process.join(timeout=remaining) call
• preserve the existing deadline, atomic-file response read, timeout, terminate/kill/reap, and fallback behavior

Validation

• focused timeout/transport/fallback/route canary: 13 passed in 12.94s (CPU0)
• three uninstrumented graph-command pairs on CPU0: mean wall 13.30s -> 10.95s (-17.7%); mean user+sys CPU 11.00s -> 10.13s (-7.9%)
• git diff --check: clean
• python3 -m py_compile scripts/aoa_session_memory.py: passed with external pycache

Evidence: /srv/abyss-machine/tmp/ai/validation-speed-session-timed-join-20260907-evidence/.

The local benchmark is directional (CPU0, three pairs), not a CI or release claim. No permanent latency test was added.