A leave-p-out based estimation of the proportion of null hypotheses