Quantifying the unobserved protein-coding variants in human populations provides a roadmap for large-scale sequencing projects