Parameter-Efficient Instruction Tuning of Large Language Models For Extreme Financial Numeral Labelling