perf(backend): optimize ban lookups and prune upload group map

- Optimize the ban matching middleware by using a read-only transaction (`db.View`) for the initial scan, avoiding the single bbolt write lock on every request when no ban matches.
- Implement periodic pruning of stale entries in the upload grouper map to prevent unbounded memory growth over time.
- Avoid redundant parsing of the `max_days` form value in the upload handler.

backend/libs/handlers/upload_group.go

@@ -16,13 +16,18 @@ const uploadGroupWindow = 20 * time.Second
 // value (per account/IP) within uploadGroupWindow are grouped into one box.
 const uploadBatchHeader = "X-Warpbox-Batch"
 
+// uploadGroupPruneInterval is how often entryFor drops stale entries so the map
+// can't grow without bound (one key per account/IP + batch value otherwise).
+const uploadGroupPruneInterval = 5 * time.Minute
+
 // uploadGrouper tracks the most recent box per batch key so opt-in batched
 // uploads land in a single box. Each key has its own lock, which also serialises
 // that key's concurrent uploads so they append to the same box instead of racing
 // to create several.
 type uploadGrouper struct {
-	mu      sync.Mutex
-	entries map[string]*uploadGroupEntry
+	mu        sync.Mutex
+	entries   map[string]*uploadGroupEntry
+	lastPrune time.Time
 }
 
 type uploadGroupEntry struct {
@@ -40,6 +45,7 @@ func newUploadGrouper() *uploadGrouper {
 func (g *uploadGrouper) entryFor(key string) *uploadGroupEntry {
 	g.mu.Lock()
 	defer g.mu.Unlock()
+	g.pruneLocked(time.Now())
 	entry, ok := g.entries[key]
 	if !ok {
 		entry = &uploadGroupEntry{}
@@ -47,3 +53,24 @@ func (g *uploadGrouper) entryFor(key string) *uploadGroupEntry {
 	}
 	return entry
 }
+
+// pruneLocked drops entries whose last use is well past the grouping window so
+// the map stays bounded to recently-active keys. Callers must hold g.mu. Entries
+// currently in use, or freshly created but not yet used (zero timestamp), are
+// kept to avoid removing one a request is about to populate.
+func (g *uploadGrouper) pruneLocked(now time.Time) {
+	if now.Sub(g.lastPrune) < uploadGroupPruneInterval {
+		return
+	}
+	g.lastPrune = now
+	for key, entry := range g.entries {
+		if !entry.mu.TryLock() {
+			continue
+		}
+		stale := !entry.at.IsZero() && now.Sub(entry.at) > 2*uploadGroupWindow
+		entry.mu.Unlock()
+		if stale {
+			delete(g.entries, key)
+		}
+	}
+}