Smarter keys

[teunbrand]

This PR aims to fix #3648. This PR is mutually incompatible with #3649.

Briefly, it automatically checks whether the key value is actually in the data and doesn't draw keys when it isn't. This contrasts with the approach taken in #3649, where the user is given more control.

Less briefly: it performs this check when show.legend = NA, which is the default. When show.legend = TRUE, is draws the keys regardless of presence in the layer data. To get back the old behaviour and you have multiple legends, you'd have to be specific to include a layer in the keys of a particular legend, for example show.legend = c(colour = TRUE).

A small demo from #3648 (comment). Note how there is no point displayed for the 'regression line' key, and no line for the 'points' key.

devtools::load_all("~/packages/ggplot2/")
#> ℹ Loading ggplot2

ggplot(mtcars, aes(disp, mpg)) + 
  geom_point(aes(color = "points")) +
  geom_smooth(aes(color = "regression line"))
#> `geom_smooth()` using method = 'loess' and formula = 'y ~ x'

^{Created on 2023-05-09 with reprex v2.0.2}

[teunbrand]

Checking the data has to happen before the defaults are filled in and after_scale() modifications are applied. Otherwise the information might be lost. I gave the scales list a method to check the data. The downside is that we have to check all the data regardless of whether the scale has a legend guide, as that isn't resolved at this point.

[teunbrand]

At this point, member is a length(pal) * length(data) logical matrix tracking whether a palette value was seen.

[teunbrand]

Also aes can be a vector, e.g. c("colour", "fill"). The inclusion/exclusion rule in the guide only needs to match one of the aesthetics, not all, to be included.

[teunbrand]

Closing this PR in favour of #5502