do not unnecessarily return a union in extract

[tiemvanderdeure]

when extracting with skipmissing=false, the geometry type will currently always be a Union{Missing, ...}, even if none of the provided geometries were missing.

This PR makes it so that the type of the geometries in input and output will match.

I had to change _rowtype to do this, and changed it a little more to make it easier to reuse (which I want to do for Rasters.sample).

I also added some missing text to the docstring.

[tiemvanderdeure]

Not 100% sure if this is a bugfix or a slightly breaking change, though.

[tiemvanderdeure]

Just copied the _rowtype changes to this branch so we can figure out type stability in this PR and how Rasters.sample should work in the other.

[tiemvanderdeure]

Just to explain this latest optimization (and I don't know if this is the optimal way to go about it). Before this PR we would convert whatever we extracted from the Raster/RasterStack in a for loop, inside _maybe_add_fields. This generates an allocation for every single point that is extracted.

After this PR the point is first extracted and stored in a Vector with the same type (as if skipmissing = false), then this whole vector is converted to the skipmissing type with a broadcast. This generates one single allocation.

MWE:

dimz = (X(9.0:1.0:10.0), Y(0.1:0.1:0.2))
rast = Raster(Union{Int,Missing}[1 2; 3 4], dimz; name=:test, missingval=missing)
mypoints = [missing, (9.0, 0.1), (9.0, 0.2), (10.0, 0.3), (10.0, 0.2)]
mymanypoints = repeat(mypoints, 500_000)
@profview extract(rast, mymanypoints, skipmissing = true)
@time extract(rast, mymanypoints, skipmissing = true);

Before:

0.268031 seconds (3.00 M allocations: 148.774 MiB, 39.32% gc time)

After:

0.156913 seconds (13 allocations: 129.700 MiB, 7.20% gc time)

[rafaqz]

3M allocations to 13 is pretty nice!