Server: use llama_chat_apply_template to format the chat

[ngxson]

Depends on #5538 to be merged

In #5425 , I mentioned that the chat template can be (ideally) detected using model metadata tokenizer.chat_template, but at that time, I didn't know that it is possible to access the metadata

Now that we have llama_chat_apply_template, we no longer have to worry about metadata. We can use this new function to format the chat supplied to /v1/chat/completions