Fix: lazy load embedding pipeline to eliminate cold start flash

[daveshenal]

What changed

• background.js: Extracted ensureEmbedder() helper that lazily loads
  the embedding pipeline only when actually needed. initialize() now
  checks the cache fingerprint first - on a cache hit cold start the popup
  reaches "ready" in ~20ms without touching the model at all. search()
  calls ensureEmbedder() before embedding the query, so the model loads
  on first search rather than on every worker respawn.

• popup.js: Added 150ms debounce before showing "Loading AI model…".
  If the background responds ready within that window the user never
  sees the loading pane. Only shown when there's genuinely something to wait for.

Tested

• Cold start with fresh cache → popup instant, no loading flash ✓
• First search after cold start → pipeline loads lazily (~300ms), results returned ✓
• Chrome restart → loading pane shows correctly ✓
• Stale cache / no cache → loading pane shows correctly ✓

[peplor]

LGTM