Fix systemd timeout check for missing user unit

[rodion-m]

Summary

• query LoadState together with TimeoutStopUSec in check_systemd_timing_alignment()
• skip user-scope systemd results whose unit is not-found, masked, or error before falling back to the system unit
• add regression coverage for a system service where systemctl --user show hermes-gateway returns LoadState=not-found plus the user-manager default timeout

Why

On a system-level Hermes gateway deployment, the current check tries systemctl --user show hermes-gateway --property=TimeoutStopUSec first. On Ubuntu/systemd, that can return success for a missing user-scope unit with LoadState=not-found and the manager default TimeoutStopUSec=1min 30s. Hermes then logs a false stale-unit warning even when the real system service has an aligned timeout.

Observed on a live v0.18.0 system-service gateway:

• user-scope query: LoadState=not-found, TimeoutStopUSec=1min 30s
• system query: effective timeout >= drain timeout + headroom
• warning before this patch: Stale systemd unit detected: hermes-gateway.service has TimeoutStopSec=90s...

Fixes #36755.

Test

• pytest tests/gateway/test_shutdown_forensics.py -q

[alt-glitch]

This was generated by AI during triage.

Duplicate of #34734 — same mechanism (query LoadState alongside TimeoutStopUSec in check_systemd_timing_alignment() and skip user-scope units whose LoadState is not-found/masked/error). #34734 is the earliest open PR using this approach; #36766 and #37324 are the same fix. #54396 solves the same #36755 false-positive via a different mechanism (detect the systemd manager from /proc/self/cgroup) and is related rather than a duplicate. Flagging the cluster so a maintainer can pick one.

[rodion-m]

Closing as duplicate after triage. This fix belongs to the existing upstream cluster for #36755; #34734 is the earliest open PR using the same LoadState-based approach, and #37324 carries the same approach with regression coverage. Our production verification still confirms the bug is present in v0.18.0, but keeping another duplicate PR open would add noise for maintainers.