Quantifying the Ease of Reproducing Training Data in Unconditional Diffusion Models