stack invalid values as NaN

[Fil]

We have originally stacked invalid values as 0 height segments (where X1[i] == X2[i] is the current height of the stack). But I think they should be ignored instead (X1[i] = X2[i] = NaN). This feels more principled.

Pros:

• cleaner information (an invalid value is kept throughout)
• allows to draw interrupted stacked lines

Con:

• gaps appear when an area "suddenly disappears" instead of vanishing in a continuous way (but this is "wrong", and can be solved by mapping NaNs to zeroes if the user wants, so it's not removing any feature).

before	after

blocks #2133

[mbostock]

We can trace this design back to d3/d3-shape#52 (2015!) where d3.stack treats NaN lengths as zero, rather than propagating the NaN to the stacked extent.

I found another case where a zero was creeping in (the emptyFacet test, where the VALUE column is undefined), and simplified the logic.

[mbostock]

Previously: #2178 #2179.