Agent LXC filesystem collection starves later containers because every pct df shares one 10s budget #1784

Problem

Reported by @avsdev-cw in #1477 (comment). On hosts with many running containers only the first handful report per-mount filesystem usage; the rest fail with signal: killed in the agent debug log. pct df takes the container config lock and costs over a second per guest, and every per-container query shared one 10-second budget with pct list in internal/hostagent/proxmox_lxc_filesystems.go, so the deadline expired after roughly five containers and the remaining queries were killed mid-run.

Fix on main

Commit 8318c74:

• resolves per-mount usage with statfs through /proc/<pid>/root (config-declared mounts read from /etc/pve/lxc, init PID via lxc-info), which takes no config lock and needs no per-guest subprocess
• drops config mounts whose device identity matches their parent directory instead of reporting the parent filesystem's numbers for an unmounted path
• keeps pct df as a per-container fallback for least-privilege installs that cannot traverse /proc/<pid>/root, now with a per-container deadline inside a bounded collection window so one slow guest cannot starve the rest
• logs a warning naming how many containers were skipped if the collection window is ever exhausted, instead of failing silently at debug level
• regression coverage pins the statfs path being preferred over pct df, per-container fallback isolation, and config parsing stopping at snapshot sections

Retest boundary

Keep open until a release containing 8318c74 ships and the reported many-container host shows filesystem usage for all running containers (up to the 128-container inventory cap).

_{How Pulse handles triage}