Prefilling assistant message in openai compatible API

[matteoserva]

This adds support for prefilling assistant response (or its thought process) using the OpenAI compatible API.

The feature is used for example by Claude.

It can be tested using open-webui or with the following curl command:

curl http://localhost:8080/apply-template \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"model": "gpt-3.5-turbo",
"messages": [
 {
    "role": "system",
    "content": "SYSTEM"
 },
 {
    "role": "user",
    "content": "USERMESSAGE"
 },
 {
    "role": "assistant",
    "content": "ASSISTANT"
 }
]
}'

Example advanced scenario: time limit for the thinking process

• launch a reasoning model and stop its thought early
• append </think> to its partial response
• prefill the response and let it continue generating tokens

[isaac-mcfadyen]

Just a heads-up that this is potentially a very breaking change, especially because this is an OpenAI compatible API but this is not OpenAI's behavior.

The main situation I can think of is if someone wants to generate a new assistant message after the last one - i.e for ChatML they want the <|im_end|><|im_start|>assistant added between the last message and the new one, rather than the last message to just be continued.

I'd suggest we add this to #9291 at a minimum.