Add a "full series" option to the stack transforms

[Fil]

See discussion at #325 and #348 (comment)

[mbostock]

Datadog calls this “default zero” interpolation: https://docs.datadoghq.com/dashboards/functions/interpolation/#default-zero

I wonder to what degree this is specific to time series. I can certainly imagine cases where it’s not specific to time series, but when it is, it seems like the bin transform with filter: null is an option for fixing the missing data. Edit: Okay, the example histogram you made is pretty convincing that we shouldn’t think of this as only a time-series problem. (Also in a related irony, this Cloud Costs notebook demonstrates the problem, but has another problem of time being represented as ordinal strings.)

[Fil]

It's worse than this: using the empty bin approach is necessary for a continuous ("binnable") domain, but far from sufficient—as soon as you have z or facets, you need a point (real or fake) for each element of the domain times each of the series.

[mbostock]

But the bin domain is the same across all groups and facets, so as long as you have at least one data point in a given group, you’ll get all the bins?

[Fil]

In https://observablehq.com/d/f6a7975f2ad4519a there is just one empty bin (4,750 in in the chinstrap facet), which should be mapped to 0 for data fidelity. If we push up the number of bins to 200, we start to see that issue creeping everywhere — outlined in red in the image below, all the areas should drop to zero since there is no data point in this position.

(I'm not sure that we can find a generic way to do both operations, maybe imputing missing values is something that should be left to the data-wrangling section?)

[mbostock]

But that example doesn’t use filter: null on the bin transform, right?

[Fil]

the Plot.seriesX transform is possibly superseded by #499 and #500

[mbostock]

I think this is probably now a duplicate of #597. (And perhaps #513 in the case of ordinal data.) At any rate they’re all closely related.