Port Jingjing's backcasting preprocessing utils

[brookslogan]

Resolves #88. Might involve #49, #90, #106, #109.

Progress so far:

• Ported over some of Jingjing's preprocessing functions and tests.
• Got the tests running in a package environment.

Some TODOs:

• Get these working based on epiprocess classes (taking care that current functions work on a single geo/epigroup at a time)
• Work either on implied lag version - time_value or an explicit lag/lag-like column.
• Check that, or change to: either always, or optionally, add NAs getting/filling lag training data if it looks like we missed recording a version (it's okay if there is no update for an observation in a target version, but a problem if there are no updates for any observations in a target version) unless there are bugs from assuming that versions are evenly spaced, this is a separate convenience function/arg to think about independently
• Change fill_rows and fill_missing_updates to use a mix of last-version-carried-forward and NA/0/customizable fill-in determined by archive's [check out $fill_through_version --- may or may not be useful]
• Add examples
• Use Abort, etc., rather than stop
• Think about interaction with epix_slide. Straightforward is probably combining with Should slide() for epi_archive be given access to less than the most up-to-date snapshots? #49. But we might also think about turning the time&version-lag covariates dfs into a custom type of epi_archive (base epi_archive would be pretty inefficient though; lazy merge might be one way to improve). ------ Just do the straightforward way to begin with, then profile.
• [Use alternative to zoo::rollmeanr. Maybe data.table::frollmean?]
• [Check compatibility of "shift" terminology with tidymodels, which is used by epipredict. (Might be incompatible with some of our production COVID-19 hospitalization forecasters, but that might be fine.)]