Consider sorting instead of hashing for uniqueness

[hadley]

https://news.ycombinator.com/item?id=21985246

This would allow us to use radix sort of for integer indices, which would help us to match data.table performance when grouping in dplyr.

[DavisVaughan]

See #1361 (comment) for explanation of why we probably don't want switch all of the dictionary functions over to sorting

We did expose vec_locate_sorted_groups() (which uses radix ordering) for use in dplyr's group_by() tidyverse/dplyr#6297, which was really the original motivation of this issue