refactor rotary embeddings

Author: DavidLandup0

keras_hub/src/models/smollm3/smollm3_causal_lm.py

@@ -70,7 +70,7 @@ def call_with_cache(
         x = self.backbone.token_embedding(token_ids)
 
         # Each decoder layer has a cache; we update them separately.
-        position_embeddings = self.backbone.rotary_embedding(x)
+        position_embeddings = self.backbone.rotary_embedding(x, start_index=cache_update_index)
         updated_cache = []
         for i in range(self.backbone.num_layers):
             current_cache = cache[:, i, ...]