Improve mapping bias estimates

[lima1]

Current heuristic simply uses neighboring SNPs. Would be better to train a regression model using database generating that takes distance, number of samples and coverage into account. Simple to benchmark by leave-one-out simulations.