`drop_na()` with an empty selection shouldn't be the same as no selection

[DavisVaughan]

library(tidyr)

df <- data.frame(x=c(1, NA, 2),y=c(NA, 1, 2))

# People like to use drop_na() like this
# (uses all columns)
drop_na(df)
#>   x y
#> 1 2 2

# But this currently also resolves to "use all columns"
# when really it should be "don't use any columns"
drop_na(df, starts_with("foo"))
#>   x y
#> 1 2 2

^{Created on 2021-11-19 by the reprex package (v2.0.1)}