Assisstant prefill support

"messages": [

{

"role": "assistant",

"content": "",

"reasoning_content": "The reasoning in the response will start with this"

}

],

When messages end with an assistant message, it should be used to prefill the generated response, which is standard behavior for self hosted llama.cpp.