Personalisation within bounds: A risk taxonomy and policy framework for the alignment of large language models with personalised feedback