consider replacing masked pixels with model

[esheldon]

We already do this for the EM algorithm, we could do it for admom and maybe others as well.

We might have some evidence this can provide additional stability from comparisons with HSM