Fix across() and cur_data() to work sequentially

[DavisVaughan]

Closes #4907

This PR fixes both across() and cur_data() when used in sequential expressions.

This was actually incredibly hard to get right. This is complicated by the fact that we don't want to resolve "unused" columns, which would mark them as "used", but we still have to check their types in eval_select() if something like across(is.numeric) is called.

The idea here is to separate the non-group columns that are currently present in the data into two groups. The first group are unused columns, which we pull from the original data. The second are used columns, which we pull from the current bindings.

The good news is that, even though this is more expensive, it is part of the "set up" work that gets cached in #4909. I'll rework that PR on top of this

[DavisVaughan]

I also discovered to my horror that if you c() two length-0 named lists together, you get an unnamed list back. This broke eval_select() (as it should), but luckily vec_c() works correctly.

library(vctrs)
library(tidyselect)
library(rlang)

x <- list(x = 1)
x <- x[0]
x
#> named list()

# WHY?
c(x, x)
#> list()

vec_c(x, x)
#> named list()

eval_select(expr(everything()), c(x, x))
#> Error: Can't select within an unnamed vector.

eval_select(expr(everything()), vec_c(x, x))
#> named integer(0)

^{Created on 2020-02-28 by the reprex package (v0.3.0)}