TASK-032: Implement geo_cache retention policy and cleanup

Add automatic cleanup of stale geolocation cache entries to prevent
unbounded database growth. Resolves the issue where unique IP addresses
accumulated indefinitely in the geo_cache table, degrading query performance.

## Changes

### Database Schema (Migration 3)
- Add 'last_seen' column to geo_cache table tracking last reference time
- Existing entries default to current timestamp

### Repository Layer (geo_cache_repo.py)
- Update upsert_entry() to set/refresh last_seen on insert/update
- Update upsert_neg_entry() to set/refresh last_seen on negative cache hits
- Update bulk_upsert_entries() to set/refresh last_seen in batch operations
- Add delete_stale_entries(db, cutoff_iso) -> int for purging old entries

### Background Task (geo_cache_cleanup.py)
- New APScheduler task that runs nightly (24-hour interval)
- Calculates cutoff as 90 days ago from current time (UTC)
- Deletes all entries with last_seen older than cutoff
- Logs operation results (info when deleted > 0, debug when 0 deleted)
- Configurable retention period via GEO_CACHE_RETENTION_DAYS constant

### Application Startup (startup.py)
- Register geo_cache_cleanup task in scheduler during app startup
- Placed after geo_cache_flush in task registration order

### Tests
- Add delete_stale_entries test cases covering:
  * Removal of old entries beyond cutoff
  * No deletion when all entries are recent
  * Empty table edge case
- Update existing test fixtures to include last_seen column
- Add full test suite for cleanup task registration and execution

### Documentation
- Architekture.md: Document cleanup task, update schema/diagram
- Backend-Development.md: Add retention policy documentation

## Behavior

When an IP is accessed, its last_seen is refreshed. After 90 days of no
access, an IP is purged by the nightly cleanup. On next encounter, the IP
is re-resolved from MaxMind MMDB or ip-api.com (if configured).

This is acceptable because:
1. Stale geolocation data may become inaccurate over time
2. Re-resolution cost is minimal compared to unbounded storage growth
3. Active IPs maintain fresh data through their last_seen updates

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

backend/app/startup.py

@@ -19,7 +19,15 @@ from apscheduler.schedulers.asyncio import AsyncIOScheduler  # type: ignore[impo
 from app.db import init_db, open_db
 from app.services import setup_service
 from app.services.geo_cache import GeoCache
-from app.tasks import blocklist_import, geo_cache_flush, geo_re_resolve, health_check, history_sync, session_cleanup
+from app.tasks import (
+    blocklist_import,
+    geo_cache_cleanup,
+    geo_cache_flush,
+    geo_re_resolve,
+    health_check,
+    history_sync,
+    session_cleanup,
+)
 from app.utils.async_utils import run_blocking
 from app.utils.jail_config import ensure_jail_configs
 from app.utils.runtime_state import set_runtime_settings
@@ -185,6 +193,7 @@ async def startup_shared_resources(
 
         health_check.register(app)
         await blocklist_import.register(app)
+        geo_cache_cleanup.register(app)
         geo_cache_flush.register(app)
         geo_re_resolve.register(app)
         history_sync.register(app)