make legend detection code more robust.

[clauswilke]

Fixes #3963.

library(ggplot2)

nc <- sf::read_sf(system.file("shape/nc.shp", package="sf"))
nc$BIR74_bin <- cut_number(nc$BIR74, 3)
nc_tibble <- tibble::as_tibble(nc)

ggplot(nc_tibble) +
  geom_sf(aes(geometry = geometry, fill = BIR74_bin))

^{Created on 2020-04-26 by the reprex package (v0.3.0)}

[yutannihilation]

For GitHub CI failures:

• I don't know what happens on Windows runner, but I bet it's just not ready for R 4.0.
• For Linux, something goes wrong but I have no idea... (AFAIK, libcurl4-gnutls-dev conflicts with libcurl4-openssl-dev, but why does this conflict arises at this timing...?)

The following packages have unmet dependencies:
 libgit2-dev : Depends: libcurl4-gnutls-dev but it is not going to be installed
E: Unable to correct problems, you have held broken packages.

Anyway, either of these is our fault, so you can just ignore the failures for now.

[yutannihilation]

Looks good! Could you add a NEWS item?

[clauswilke]

What exactly do you think the NEWS item should say? Just something like "fixed a bug in autodetection of legend type for geom_sf()?" Or "In geom_sf(), the legend type is now detected from the mapped geometry column."?

[yutannihilation]

Since this is more of a bug fix, the former sounds appropriate to me.