Rework grid functions to be more robust to duplicate auto names

[DavisVaughan]

Related to auto-naming:
Closes #1221
Closes #1116
Closes #1092
Closes #1037
Closes #992

Other:
Closes #1189

This was a tricky one to fix.

Ultimately, I combined a few ideas together:

• We essentially use quos_auto_name(repair_auto = "unique", repair_quiet = TRUE) from Add repair_auto and repair_quiet to exprs_auto_name() r-lib/rlang#1194, but that is in rlang 1.0.0 and we don't use dev rlang yet. So I've just inlined the idea of it for now. Notably, that only performs unique repair on the auto-named inputs, which is definitely the right approach. This avoids getting duplicate names like <int> or truncated names like tibble(...) when a long tibble expression is inlined.
• We take an approach more similar to tibble:::tibble_quos() for iterative expressions (i.e. when the user does x = 1, y = x). This involves setting up a data mask that is separate from the actual output object.
• Unnamed data frames are given a list element name of "" before returning from grid_dots(), which serves as a signal that they should be unpacked when supplied to data_frame(). This means we were able to get rid of all usage of flatten_at() and flatten_nested().

No changes to worse in revdeps