Control for false discovery instead of padding p-values

[axiomcura]

In my opinion the non-sig decision shouldn't need padding and we should be safe to assign all features whose adj p > 0.05 as non significant. The ambiguous features would be the band that falls within (p_threshold - padding, p_threshold) , and everything from (0, p_threshold - padding] would be the significant as you are already doing. Ideally we would want a data driven way to determine the band width to avoid dropping too many features. One simple way of doing this is to set a threshold for the max allowed false discovery feature count (e.g. 5) and treat the adj p values themselves as the probability of false positive, and you choose the smallest bandwidth/padding that results in the sum(adj p of all features within the band) < false_discovery_feature_count. This also makes this argument more interpretable.

Originally posted by @wli51 in #39 (comment)