fix(llm): cap auto-detected max_output_tokens when it fills the entire context window

fix(llm): cap auto-detected max_output_tokens when it fills the entire context window #12436