Eliminate/explicate differences in training windowing between flatline and arx forecasters

[brookslogan]

#290 highlighted that training window sizes similar to the ahead value can trip up the flatline forecaster. But this also indicates that the flatline forecaster is not using anywhere near n_training instances per epikey if ahead is within an order of magnitude of n_training. This is not the case for arx_forecaster:

library(epipredict)
#> Loading required package: epiprocess
#> 
#> Attaching package: 'epiprocess'
#> The following object is masked from 'package:stats':
#> 
#>     filter
#> Loading required package: parsnip
trace(slather, quote({
  if (inherits(object, "layer_residual_quantiles")) {
    trace(dplyr::summarize, quote({
      cat("Number of non-NA residuals:\n")
      print(.data %>% tidyr::drop_na(.resid) %>% nrow())
    }))
  }
}), quote(untrace(dplyr::summarize)))
#> Tracing function "slather" in package "epipredict"
#> [1] "slather"
case_death_rate_subset %>% flatline_forecaster("case_rate", flatline_args_list(ahead = 28L, n_training = 29L))
#> [...]
#> Number of non-NA residuals:
#> [1] 56
#> [...]
case_death_rate_subset %>% arx_forecaster("case_rate", "case_rate", args_list = arx_args_list(ahead = 28L, n_training = 29L))
#> [...]
#> Number of non-NA residuals:
#> [1] 1624
#> [...]

^{Created on 2024-04-19 with reprex v2.0.2}

However, ?flatline_args_list doesn't explicate this

n_training: Integer. An upper limit for the number of rows per key that
          are used for training (in the time unit of the 'epi_df').

and the message from slather.layer_residual_quantiles when output residuals are NA is something specific to flatline forecaster (and off by one for flatline_forecaster):

! Residual quantiles could not be calculated due to missing residuals.
ℹ This may be due to `n_train` < `ahead` in your <epi_recipe>.

Approach 1: eliminate these differences. Make n_training make sense for flatline_forecaster by using the same NA omission pre training window approach as arx_forecaster. Remove the mention of the inequality above in the layer_residual_quantiles error message since it won't be an issue anymore.

Approach 2: explain the difference in ?flatline_args_list, and mention n_train --> <= <-- ahead is an issue --> for flatline_forecaster <-- in the residual quantiles error message.