`absorb=TRUE` as column transformations coded in a model formula instead of transformed columns added to the original dataframe

When `absorb=TRUE`, we block-center treatment and moderator variables. Without a moderator or with a categorical moderator, these transformations add new columns to the `data` argument of `lmitt()`, but for continuous moderator variables, the moderator is centered in place, that is, the original variable is replaced by the block-centered variable. It would simplify `cov_adj()` if block-centering was part of the formula rather than part of the data.

We could just rename the block-centered moderator variables and alter the formula to reference them, but I think it's cleaner to address it in the formula. This blocks the way I intend to address #244