[Core] More fixes to MultiModalEmbeddings type handling (#19715)

Signed-off-by: Russell Bryant <rbryant@redhat.com>

vllm/model_executor/models/granite_speech.py

@@ -721,7 +721,8 @@ class GraniteSpeechForConditionalGeneration(
         multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
     ) -> torch.Tensor:
         """Compute the merged LLM / audio embeddings."""
-        if multimodal_embeddings is None:
+        if multimodal_embeddings is None \
+            or len(multimodal_embeddings) == 0:
             return self.language_model.get_input_embeddings(input_ids)
 
         inputs_embeds = embed_multimodal(