feat(metrics): collect host Pressure Stall Information (PSI)

[123123213weqw]

What

Adds a host Pressure Stall Information (PSI) collector reading /proc/pressure/{cpu,memory,io} and exposing some/full avg10/avg60/avg300 gauges and some/full total counters, labeled by resource.

Closes #384.

Why

Memory, CPU, and I/O contention stalls are the earliest signal of resource pressure, but huatuo had no PSI collection — operators only see the fallout (OOM, request timeout, node stall) after the fact. PSI directly quantifies how long tasks were stalled by resource contention.

How

• New core/metrics/system_psi.go collector (system_psi), following the loadavg / cpu_stat host-metric pattern (FlagMetric, polled via Update()).
• Reads /proc/pressure/{cpu,memory,io} via the vendored procfs.FS.PSIStatsForResource (re-exported PSIStats/PSILine type aliases from internal/procfs, alongside the existing FS/ProcMap aliases).
• Emits, per resource: psi_some_avg10/avg60/avg300, psi_full_avg10/avg60/avg300 (gauges) and psi_some_total, psi_full_total (counters), labeled resource=cpu|memory|io. cpu has no full line (kernel limitation), so it produces only the some series.
• Graceful degradation: if /proc/pressure is absent (PSI disabled at boot) the collector returns types.ErrNotSupported at registration → marked inactive, no impact on other collectors; per-resource read failures are skipped rather than failing the scrape.

Changes

• core/metrics/system_psi.go (new) — collector + psiSamples mapping.
• core/metrics/system_psi_test.go (new) — psiSamples cases (cpu some-only, memory some+full, empty stats).
• internal/procfs/fs.go — re-export PSIStats, PSILine type aliases.

Verification

Built and tested on the wzu host (Linux 6.8, /proc/pressure live):

• ✅ go build ./...
• ✅ go test ./core/metrics/... ./internal/procfs/... (incl. new TestPSISamples)
• ✅ go vet ./core/metrics/... ./internal/procfs/...
• ✅ goimports / gofumpt / golangci-lint (v1.62.2) clean
• ✅ make all — all 6 binaries + BPF objects build green

Out of scope (follow-ups, per the issue)

• cgroup v2 *.pressure collection + container association.
• Pressure state machine (configurable thresholds, sustained-trigger / hysteresis recovery).
• Grafana panel + alerting examples.
• Cardinality controls for high-density nodes (host PSI has only 3 resources, so not needed here).

[123123213weqw]

整理我的开放 PR 队列，先关闭。分支保留，需要时重新打开或基于最新 main 重新提交。/ Cleaning up my open PR queue — closing for now. The branch is kept; will reopen or resubmit on latest main later.