Consider allowing negative `after` values

[brookslogan]

In advanced.Rmd, we fit a nowcasting model for y given x using epi_slide. However, beyond the latency&revision-naiveness, it is also will give unrealistic performance because it includes the testing data in the training data. If we want to enable this type of operation (questionable; should we route solely to epi_archive?), it'd be nice to be able to set after = -k when the reporting latency is expected to be k.

[brookslogan]

See discussion in #288 for why after = -1 would not be appropriate here unless we separated the fit (slide with before = 2, ahead - 1) and predict (mutate/etc.) steps.

Perhaps a more important use case, motivating the addition of an after parameter and allowing it to be negative would be for epix_slide extracting a particular lag (though doing this for each lag doesn't sound great for efficiency; maybe this needs to be its own function even though technically we could do it with an epix_slide).

[dshemetov]

Closing as out of date. We no longer have the after argument.