fix: OpenAI API - together.ai extra usage chunks

This fixes an issue where, with some models (ie the Llama 4 models),
together.ai is sending a final usage chunk for streaming responses
even if the user didn't ask to include usage.

With this change, the OpenAI API verification tests now pass 100% when
using Llama Stack as your API server and together.ai as the backend
provider.

As part of this, I also cleaned up the streaming/non-streaming return
types of the `openai_chat_completion` method to keep type checking happy.

Signed-off-by: Ben Browning <bbrownin@redhat.com>

llama_stack/providers/remote/inference/ollama/ollama.py

@@ -5,7 +5,7 @@
 # the root directory of this source tree.
 
 
-from typing import Any, AsyncGenerator, Dict, List, Optional, Union
+from typing import Any, AsyncGenerator, AsyncIterator, Dict, List, Optional, Union
 
 import httpx
 from ollama import AsyncClient
@@ -41,6 +41,7 @@ from llama_stack.apis.inference import (
 )
 from llama_stack.apis.inference.inference import (
     OpenAIChatCompletion,
+    OpenAIChatCompletionChunk,
     OpenAICompletion,
     OpenAIMessageParam,
     OpenAIResponseFormatParam,
@@ -409,7 +410,7 @@ class OllamaInferenceAdapter(
         top_logprobs: Optional[int] = None,
         top_p: Optional[float] = None,
         user: Optional[str] = None,
-    ) -> OpenAIChatCompletion:
+    ) -> Union[OpenAIChatCompletion, AsyncIterator[OpenAIChatCompletionChunk]]:
         model_obj = await self._get_model(model)
         params = {
             k: v