[16.0] llm_thread refactor

LLM thread should format messages for the model before sending it to chat and control the sequence, so we need to get rid of _prepare_chat_params's system_prompt=None, prepend_messages=None as this should be controlled by the llm_thread in case of templating filtering and so on.

the model should expose format_messages() or actually we have a formatter on mail.message already

LLM thread methods need to be reviewed which are public and which are private and named consistently e.g.

call_assistant
call_tool and so