The Construction of Instruction-tuned LLMs for Finance without Instruction Data Using Continual Pretraining and Model Merging