add categorical equality fast paths

[ritchie46]

A few easy perf/memory wins.

• cat == str should first check the rev map.
• cat.is_in should first check the rev_map if is_in is reasonable sized.
• First global rev map doesn't need a map from global to local.

[alexander-beedie]

We've been moving a lot of data into Categorical cols recently, so this sounds fantastic :)

[ritchie46]

Tried the smaller rev-map, but that seems more trouble than it is worth. We cannot determine fast uniques in such a case and get regressions in string comparisons. I will go with equality fast paths.

[ghuls]

Wouldn't a bloom filter be a good alternative for a fast in set reduction approach?

[ritchie46]

Wouldn't a bloom filter be a good alternative for a fast in set reduction approach?

Yes, but we also need to map from global idx to local one. Don't want to do a linear search once the bloom filter says yes.

[ritchie46]

Will go for a btree. They are much more condensed and likely to be much faster on u32.