Track duplicates in DataAPI.refpool issue

[bkamins]

Follow up of #2442 (comment).

We should resolve it before 1.0 release as I understand that if we have duplicates in DataAPI.refpool we are going to get wrong results in split-apply-combine.

[bkamins]

CC @nalimilan @quinnj

[nalimilan]

We can fix it at any point after 1.0. Currently results are correct, we just have a check that is sometimes costly and not necessary for PooledArray and CagegoricalArray.

[bkamins]

Ah - so we have this check now? I thought we do not check uniqueness.

[bkamins]

This might be one of the reasons of slow H2O tests then for very large number of groups.

[nalimilan]

I don't think so: with 100 rows per group (instead of 5 in my example) the overhead turns out to be much smaller. I was surprised by this but it makes sense I guess. Anyway, saving a few percents wouldn't be bad.

[nalimilan]

The check is:

DataFrames.jl/src/dataframerow/utils.jl

Line 199 in 8645651

anydups = !all(allunique, refpools)