PANDORA: Diffusion Policy Learning for Dexterous Robotic Piano Playing