Fix metrics flickering and improve internals page

- Fix dashboard metrics alternating to 0 by merging partial batches
  in gateway before broadcasting to WebSocket clients. The aggregator
  sends metrics in batches of 20, causing partial updates that
  overwrote each other. Gateway now maintains machine_metrics_cache
  that accumulates metrics across batches.

- Remove misleading gRPC calls counter from internals page (only
  incremented on health checks, not actual metric flow). Replace
  with cached_machines counter showing tracked machines.

- Update internals.html stats panel to show Events, Broadcasts,
  Clients, and Machines instead of gRPC calls.

services/aggregator/storage.py

@@ -38,10 +38,21 @@ class RedisStorage:
         metrics: dict[str, float],
         timestamp_ms: int,
     ) -> None:
-        """Update the current state for a machine."""
+        """Update the current state for a machine (merges metrics, doesn't replace)."""
         if not self._client:
             raise RuntimeError("Not connected to Redis")
 
+        key = f"machine:{machine_id}"
+
+        # Get existing state to merge metrics
+        existing_data = await self._client.hget(key, "state")
+        if existing_data:
+            existing_state = json.loads(existing_data)
+            existing_metrics = existing_state.get("metrics", {})
+            # Merge new metrics into existing (new values override old)
+            existing_metrics.update(metrics)
+            metrics = existing_metrics
+
         state = {
             "machine_id": machine_id,
             "hostname": hostname,
@@ -51,7 +62,6 @@ class RedisStorage:
         }
 
         # Store as hash for efficient partial reads
-        key = f"machine:{machine_id}"
         await self._client.hset(
             key,
             mapping={