Integrate a PSI watcher to track memory and cpu pressure

To mitigate the risk of cgroups evicting filesystem cache, we can rely on the PSI metrics through which memory.pressure and cpu.pressure can be tracked. A PSI watcher would provide an earlier signal than usage thresholds. For instance, Meta's oomd uses PSI with a 3-condition check:

• Sustained 60s pressure above threshold
• Current 10s elevated pressure
• Pressure is not already recovering.

These conditions avoid false positives from brief spikes. Once these thresholds are crossed, we can emit backoff events.