feat: make embedding model configurable for semantic search

- Add SemanticSearchConfig to settings.py with embedding model options
- Update WorldMapper to accept embedding_model parameter
- Update main.py to pass model from config to WorldMapper
- Add documentation for available embedding models:
  - sentence-transformers/all-MiniLM-L6-v2 (default, fast)
  - sentence-transformers/all-mpnet-base-v2 (better quality)
  - google/embeddinggemma-300m (state-of-the-art)
- Update config.yaml and config.example.yaml with new options

Users can now choose the best embedding model for their needs.

examples/config.example.yaml

@@ -35,10 +35,15 @@ mapper:
       fps: 30
       codec: "h264"
   
-  semantic_search:
-    model: "all-MiniLM-L6-v2"
-    top_k: 5
-    similarity_threshold: 0.7
+  semantic_search:
+    # Embedding model for semantic room search
+    # Options:
+    #   - "sentence-transformers/all-MiniLM-L6-v2" (default, fast, 22MB)
+    #   - "sentence-transformers/all-mpnet-base-v2" (better quality, 418MB)
+    #   - "google/embeddinggemma-300m" (state-of-the-art, requires transformers>=4.40)
+    model: "sentence-transformers/all-MiniLM-L6-v2"
+    top_k: 5
+    similarity_threshold: 0.7
 
 behavior:
   reaction_time_min_ms: 200