Clarify documentation for and sanity-check nafill_buffer usage

[brookslogan]

The args list help entries for the canned forecasters, including ones that don't use lags, is:

nafill_buffer: At predict time, recent values of the training data are
          used to create a forecast. However, these can be 'NA' due to,
          e.g., data latency issues. By default, any missing values
          will get filled with less recent data. Setting this value to
          'NULL' will result in 1 extra recent row (beyond those
          required for lag creation) to be used. Note that we require
          at least 'min(lags)' rows of recent data per 'geo_value' to
          create a prediction. For this reason, setting 'nafill_buffer
          < min(lags)' will be treated as _additional_ allowed recent
          data rather than the total amount of recent data to examine.

This confused me, and I think there may be some implementation bug(s).

Requests:

• State the default value, Inf, and what it does.
• State what "regular" values (non-NULL, not < min(lags)) do.
• "Setting this value to 'NULL' will result in 1 extra recent row (beyond those required for lag creation) to be used." ---
  • An extra row from where, for what? --- I think I understand now after a while of thinking about the name and looking at the implementation, but things could be clearer.
  • Why would this be expected to work so much that it's a special value?
  • It looks like "those required for lag creation" corresponds to max_lags + max_horizon, but factors into keep in x <- dplyr::filter(x, forecast_date - time_value <= keep), which
    • if forecast_date is the default max time_value, seems like more than is required for lag creation.
    • if forecast_date has been passed as the actual forecast date, makes it seem that max_horizon should be something around the reporting latency instead, although maybe this is off by one since keep is at minimum min_required + 1 not min_required.
    • (Also, this type of filter calls into question the naming n_recent.)
  • Actually, trying to use this, I get

out <- arx_forecaster(
  case_death_rate_subset, "death_rate",
  "death_rate",
  args_list = arx_args_list(
    ahead = 16,
    lags = 7,
    n_training = 30,
    nafill_buffer = NULL
  )
)
#> Error in if (is.finite(nafill_buffer)) arg_is_pos_int(nafill_buffer, allow_null = TRUE) : 
#>   argument is of length zero

• ("For this reason, setting 'nafill_buffer < min(lags)' will be treated as additional allowed recent data rather than the total amount of recent data to examine" --- still sounds like a footgun to have < min(lags) and >= min(lags) do different things.)
• Note/remind that this NA filling is not used in the training process?
• For forecasters that don't use lags, provide an explanation that doesn't mention lags?

[brookslogan]

Maybe we don't have to address this, if the related code and documentation is removed as part of #293.