Improve perforamance of _pluck_uniq_cols

[mlazowik]

Use a set instead of an OrderedDict with empty values for gathering
unique values.

O(n^2) -> O(nlogn)

[coveralls]

Coverage decreased (-0.7%) to 83.255% when pulling b37f9e6 on mlazowik:perf-set into 8628d7e on kobotoolbox:master.

[mlazowik]

This buys around 2-3s per reports/{uid} request on a form with ~2k questions and ~100 revisions.

[dorey]

Thanks, @mlazowik for looking into this. It looks like a clear improvement

[dorey]

I think this was good at db116cc . It was clear what the code was doing and seemed to be an improvement.

I'd rather merge that than include the orderedset library unless it's helpful for further performance improvements.

Please note, that a lot of the formpack code will be in flux in the next month as we move asset.content in kobotoolbox/kpi over to a json-schema validated structure, which opens up the possibility of loading in just the diffs between versions.

[mlazowik]

Right, I added the dep b/c I wanted to extract what I've written into utils/ordered_set.py so that it's not loose in a random place, but then I noticed there is already a library for that.

I didn't see any more perf. improvement over my code. I can revert the dep, what do you think about extracting my little code to utils/ordered_set.py?

[dorey]

If it were going to be reused elsewhere in the code I think that would be helpful. The person who helped us early on was fond of using OrderedDict() with empty values as a workaround for not having a native python implementation of ordered set. I was not aware of performance cost of this approach.