fix: add default TimeThresholds to prevent hair-trigger alerts (fixes #491)

Root cause: NewManager() was missing TimeThresholds initialization, causing
all alert types to use 0-second delay. This meant alerts fired immediately
on the first sample exceeding threshold, with no debouncing.

Impact: LXC containers with brief CPU spikes to ~100% (normal for single-core
saturation) triggered constant alerts instead of only alerting on sustained
high CPU usage.

Fix: Add default TimeThresholds:
- guest: 10s delay (prevents alerts from brief CPU spikes)
- node: 15s delay
- storage: 30s delay
- pbs: 30s delay

This ensures CPU must stay above threshold for the configured duration
before an alert fires, preventing noise from momentary spikes.

Fixes #491

internal/alerts/alerts.go

@@ -247,7 +247,13 @@ func NewManager() *Manager {
 			MinimumDelta:      2.0, // 2% minimum change
 			SuppressionWindow: 5,   // 5 minutes
 			HysteresisMargin:  5.0, // 5% default margin
-			Overrides:         make(map[string]ThresholdConfig),
+			TimeThresholds: map[string]int{
+				"guest":   10, // 10 second delay for guest CPU alerts
+				"node":    15, // 15 second delay for node alerts
+				"storage": 30, // 30 second delay for storage alerts
+				"pbs":     30, // 30 second delay for PBS alerts
+			},
+			Overrides: make(map[string]ThresholdConfig),
 			Schedule: ScheduleConfig{
 				QuietHours: QuietHours{
 					Enabled:  false, // OFF - users should opt-in to quiet hours