DEFT: Efficient Finetuning of Conditional Diffusion Models by Learning the Generalised $h$-transform