`fct_unique()` improvements

[hadley]

• Make purpose more clear in docs. Fixes #284.
• Include implicit NA if present. Fixes #293.

[hadley]

From @DavisVaughan:

For future readers, we discovered that while most factors replace NA with an integer code when NA is used as an explicit factor level, it is actually documented in ?factor that is.na<-.factor can be used to simultaneously have NA as a code and NA as a completely separate explicit factor level that maps to an integer code.

x <- factor(c("a", NA, NA, "b"), levels = c("a", "b", NA), exclude = NULL)
unclass(x)
#> [1] 1 3 3 2
#> attr(,"levels")
#> [1] "a" "b" NA

is.na(x)[3] <- TRUE

x
#> [1] a    <NA> <NA> b   
#> Levels: a b <NA>
unclass(x) # technically 3 and NA are different
#> [1]  1  3 NA  2
#> attr(,"levels")
#> [1] "a" "b" NA

So this test ensures we support that