`cov_adj()` support for `medley` objects combining multiple model fits

[benthestatistician]

When there are missing covariates, a prognostic regression using those covariates as independent variables can't be used to generate response predictions. But if you had a second, backup regression that did not use those covariates as predictors, you might use the first version when it works and fall back to the second when it didn't. The medley function of the medley package automates the fitting of a given regression and one or more backups excluding covariates subject to missingness, and provides a predict() method for generating the corresponding predictions.

It would be nice to be able to accept medley objects in cov_adj()'s model=slot. Or at least those that assemble fitted models of types we can otherwise accommodate ascov_adj() model's, such as lm's, glm's, glmrobs and so forth.

This calls for planning on two levels - spelling out how to build the $A_{11}$ and $A_{21}$ matrices and the estfun object from those of the primary and backup regressions; spelling out the needed .make_PreSandwichLayer() method, and adjustments to underlying S4 object (if/as needed).

[benthestatistician]

Couple quick notes to self (for now) to maybe expand later:

• Perhaps bread and estfun methods for medley objects could be useful in contexts other than ours too. (e.g., {medley} defaults to using non-overlapping sets of observations to fit its distinct models -- if this has to do with the fact that it makes the models' coefficients uncorrelated, then perhaps there'd be a role for cross-model covariance estimates in the case where they're fit to overlapping sets of observations.)
• What does .make_PreSandwichLayer() need from its model's terms? For a medley object combining several models each of whose terms is a subset (so to speak) of the first model's, would it be enough to look at the first model's terms?
• It might be useful to be able to take a quick look at medley object to determine whether it's a medley of lm's, or of glm's or whatever. This line setting its class to medley could be readily modified to set it to c("medley", "lm"), c("medley", "glm") or whatever. Or maybe there are reasons not to do this? Perhaps better presented as an issue over there than as a PR.