Implement offline cache support in aggregation service

Implement OfflineCacheService as a thin wrapper over sqflite with a table named annual_stats_cache (columns: cache_key TEXT PRIMARY KEY, raw_stats TEXT, fetched_at INTEGER as Unix timestamp milliseconds). Store raw_stats as a JSON string (jsonEncode/jsonDecode). Use crypto package's sha256 for the cache key hash — this keeps keys at a fixed length regardless of date string format. Make the cache write fire-and-forget using unawaited(cacheService.saveAnnualStats(...)) so the fresh result is returned to the caller immediately without waiting for the disk write.

The 7-day TTL for prune is a business decision — make it a named constant (kCacheMaxAgedays = 7) in a constants file so it can be adjusted without hunting through the code. Expose cachedAt as an optional field on the AnnualStatsState (BLoC) or an AsyncValue annotation (Riverpod) rather than embedding it in AnnualStatsResult — the result model is a domain object and should not carry infrastructure concerns like cache timestamps.

Unit tests (flutter_test) with an in-memory fake OfflineCacheService. Test groups: (1) Write-through — mock repository returns success; assert saveAnnualStats was called with correct key and data; (2) Cache hit on failure — repository throws networkError; mock cache returns a CachedAnnualStats with fetchedAt = 2 hours ago; assert returned AnnualStatsResult matches cached data and cachedAt is set; (3) Cache miss on failure — repository throws networkError; cache returns null; assert AnnualStatsException is rethrown; (4) Staleness flag — CachedAnnualStats with fetchedAt = 25 hours ago has isStale == true; (5) Prune — insert entries with fetchedAt = 8 days ago and 1 day ago; call pruneExpiredEntries; assert only the 8-day-old entry is removed. Integration test: use actual sqflite in-memory database to verify the full write → read cycle round-trips correctly.