Decouple scrape path with cached scans and add scan observability

[Codex]

Prometheus scrapes were performing heavy filesystem IO inline, risking timeouts and partial data; defaults also assumed a macOS system skills path and markdown heading parsing was narrow.

• Background scanning cache
  • Added timed background scan with mutex-protected snapshot; Collect now only serves cached metrics, avoiding scrape-thread IO and tracking last duration/errors.
• Observability
  • New metrics: openclaw_scan_duration_seconds, openclaw_scan_errors_total; scan latency recorded via registered ResponseLatencyCollector.
• Robustness tweaks
  • System skills dir resolved only when explicitly set or on macOS with existing path; markdown section counting accepts whitespace after ##.

Example (cached collect path):

// Collect now reads from cached snapshot
c.mu.RLock()
snap := c.snapshot
c.mu.RUnlock()

ch <- prometheus.MustNewConstMetric(c.scrapeSuccess, prometheus.GaugeValue, snap.scrapeSuccess)