featu: support passing "extra body" throught to providers

# What does this PR do?
Allows passing through extra_body parameters to inference providers.


closes #2720

## Test Plan
CI and added new test

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -230,6 +230,9 @@ class LiteLLMOpenAIMixin(
     ) -> OpenAICompletion:
         model_obj = await self.model_store.get_model(params.model)
 
+        # Extract extra fields
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         request_params = await prepare_openai_completion_params(
             model=self.get_litellm_model_name(model_obj.provider_resource_id),
             prompt=params.prompt,
@@ -253,6 +256,7 @@ class LiteLLMOpenAIMixin(
             suffix=params.suffix,
             api_key=self.get_api_key(),
             api_base=self.api_base,
+            **extra_body,
         )
         return await litellm.atext_completion(**request_params)
 
@@ -272,6 +276,9 @@ class LiteLLMOpenAIMixin(
 
         model_obj = await self.model_store.get_model(params.model)
 
+        # Extract extra fields
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         request_params = await prepare_openai_completion_params(
             model=self.get_litellm_model_name(model_obj.provider_resource_id),
             messages=params.messages,
@@ -298,6 +305,7 @@ class LiteLLMOpenAIMixin(
             user=params.user,
             api_key=self.get_api_key(),
             api_base=self.api_base,
+            **extra_body,
         )
         return await litellm.acompletion(**request_params)
 

llama_stack/providers/utils/inference/openai_mixin.py

@@ -228,6 +228,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
         """
         Direct OpenAI completion API call.
         """
+        # Extract extra fields using Pydantic's built-in __pydantic_extra__
+        extra_body = dict(params.__pydantic_extra__ or {})
+
         # Handle parameters that are not supported by OpenAI API, but may be by the provider
         #  prompt_logprobs is supported by vLLM
         #  guided_choice is supported by vLLM
@@ -316,6 +319,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
             user=params.user,
         )
 
+        # Extract any additional provider-specific parameters using Pydantic's __pydantic_extra__
+        if extra_body := dict(params.__pydantic_extra__ or {}):
+            request_params["extra_body"] = extra_body
         resp = await self.client.chat.completions.create(**request_params)
 
         return await self._maybe_overwrite_id(resp, params.stream)  # type: ignore[no-any-return]