Fix var shifting in advanced.Rmd: by geo group, handling time gaps

[brookslogan]

Fix problem with lags not being calculated by geo group. Also edit to handle any time gaps within a geo group, or omissions at beginning or end of data set due to lag=0 not being included or covered by the range of shifts.

Current approach uses a lot of full_joins (potentially slow?) and makes prob_arx involve the separate := functionality of both rlang and data.table, which could be confusing. An alternative (below) would be using pivoting to perform the join, shown below. But this would be slightly more fragile to derive other forecasters from, as it requires the variables to have the same type, and if a variable is mapped to 0 rows, it will be silently omitted rather than filled in with NAs.

  # Build features and response for the AR model, and then fit it
  dat =
    # Make each (covariate,lag) request into a row of a tibble...
    tibble::tibble(i = seq_len(ncol(x)), lag=lags) %>%
    tidyr::unchop(lag) %>%
    # ... assign names we know will won't conflict...
    dplyr::mutate(name = paste0("x",seq_len(nrow(.)))) %>%
    # ... turn each of these requests into a separate tibble with indexing columns...
    dplyr::mutate(tbl_for_var_shift = purrr::pmap(., function(i, lag, name) {
      tibble::tibble(
                geo_value = geo_value,
                time_value = time_value + lag, # shifting values back by `lag` = shifting times forward by `lag`
                value = x[,i]
              )
    })) %>%
    dplyr::select(name, tbl_for_var_shift) %>%
    # ... as well as one for the desired shift of y...
    dplyr::bind_rows(tibble::tibble(
                               name = "y",
                               tbl_for_var_shift = list(tibble(
                                 geo_value = geo_value,
                                 time_value = time_value - ahead, # shifting variable forward by `ahead` = subtracting `lead` from time_values
                                 value = y
                               ))
                             )) %>%
    #  ... and combine them together.
    tidyr::unnest(tbl_for_var_shift) %>%
    tidyr::pivot_wider(id_cols=c(geo_value, time_value), names_from=name)

[brookslogan]

The approach to handling time gaps sort of breaks the test time selection due to creating some extra time_values that weren't in the original [and not being properly sorted by time!].

[brookslogan]

This still has time indexing issues. The problem is that if we are making a forecast at fc_time_value, we are still only getting data through, at the latest, fc_time_value-1L (could also be even earlier depending on the data sources involved). So the test_time_value does not match the fc_time_value, which is probably what we would have wanted.

Also, separately, in some debug examples, I'm seeing what looks like ~14 days of latency in the doctor-visits CLI data. I need to double-check whether I'm reading this correctly and whether it's another bug. [This latency is reflected in direct as_of queries as well. The question now is whether LOCFing over all that is all right or not for this example.]

[ryantibs]

So I just made a few small changes to the forecasters (third to last commit message gives details). Turns out turning off the intercept makes them look better overall. Will merge to main. Thanks very much @lcbrooks for the help (and for catching the bugs!!).

This closes #52 (although I only did it in the advanced vignette).