Optimise geo lookup and aggregation for 10k+ IPs

- Add persistent geo_cache SQLite table (db.py)
- Rewrite geo_service: batch API (100 IPs/call), two-tier cache,
  no caching of failed lookups so they are retried
- Pre-warm geo cache from DB on startup (main.py lifespan)
- Rewrite bans_by_country: SQL GROUP BY ip aggregation + lookup_batch
  instead of 2000-row fetch + asyncio.gather individual calls
- Pre-warm geo cache after blocklist import (blocklist_service)
- Add 300ms debounce to useMapData hook to cancel stale requests
- Add perf benchmark asserting <2s for 10k bans
- Add seed_10k_bans.py script for manual perf testing

backend/pyproject.toml

@@ -45,7 +45,7 @@ ignore = ["B008"]  # FastAPI uses function calls in default arguments (Depends)
 # sys.path manipulation before stdlib imports is intentional in test helpers
 # pytest evaluates fixture type annotations at runtime, so TC001/TC002/TC003 are false-positives
 "tests/**" = ["E402", "TC001", "TC002", "TC003"]
-"app/routers/**" = ["TC001"]  # FastAPI evaluates Depends() type aliases at runtime via get_type_hints()
+"app/routers/**" = ["TC001", "TC002"]  # FastAPI evaluates Depends() type aliases at runtime via get_type_hints()
 
 [tool.ruff.format]
 quote-style = "double"