ggdesplot(): leave cells with a missing value empty

[SchmidtPaul]

Problem

desplot() leaves a cell whose value is missing empty, ggdesplot() fills it with the
ggplot2 default grey50. That is not a neutral colour here: the default RedGrayBlue scale
runs firebrick -> lightgray -> #375997, so grey is what the middle of the data range
looks like, and a hole in the field reads as an average value rather than as no data. With a
factor response the grey cell is a fill that the key has no entry for.

Reproducer

besag.met is the dataset the vignette uses. 36 of its 1188 plots have no yield, and the gg
path drew every one of them grey (desplot 1.11, b18cfbb):

data(besag.met, package = "agridat")
f <- ggplot2::layer_data(ggdesplot(besag.met, yield ~ col * row | county), 1)$fill

sum(is.na(besag.met$yield))    #> 36
sum(f == "grey50")             #> 36    on main
sum(f == "transparent")        #> 36    with this branch

desplot(besag.met, yield ~ col * row | county)   # the same cells are empty here

The grey is hard to tell from data in that map, because the mid-range cells are themselves
grey (#D3D1D1 and #C1C5CC here).

The vignette draws this dataset with desplot() rather than ggdesplot(), so no rendered
documentation changes.

Before / after

ggdesplot(besag.met, yield ~ col * row | county), rendered against main and against this
branch. There are six missing plots in each of the six counties, so every panel is affected.

Before (main), 36 plots filled grey50:

After (this branch), the same 36 plots left empty:

Fix

na.value = "transparent" on both fill scales, scale_fill_gradientn() for a numeric
response and scale_fill_manual() for a factor. The gg panel background is
element_blank(), so a transparent cell comes out white, which is what the lattice path
draws. The observed cells are unchanged, so the fix touches the missing cells only and does
not shift the scale.

Regression test added to tests/testthat/test_ggdesplot_fixes.R. It covers both the numeric
and the factor branch on a small synthetic field, and it also pins that the remaining cells
still get a real colour, so an empty cell cannot come from a broken scale instead. On main
it fails (4 failures, the 29 existing expectations pass), with the fix the suite is 33
passing, 0 failures, 0 warnings. R CMD check --no-manual including the vignette rebuild is
Status: OK.

Prepared with the help of an AI coding assistant; the reproducers above were run and
verified locally.