Reduce backfill throttle to 0.1s to fit within gunicorn timeout

0.5s × 136 tools = 68s of sleep alone, which combined with network
time exceeded gunicorn's 120s worker timeout.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/cmdforge/registry/embeddings.py

@@ -204,7 +204,7 @@ def backfill_all_embeddings(
     ollama_url: Optional[str] = None,
     model: Optional[str] = None,
     batch_size: int = 20,
-    throttle: float = 0.5,
+    throttle: float = 0.1,
 ) -> Dict[str, int]:
     """Generate embeddings for all public/approved tools that don't have one.