Convert early_zero_weight() to data.table

[jdblischak]

Please merge PR #143 first. I will likely need to resolve some merge conflicts locally before we can merge this one

I also added some more tests

The speed increase for the unstratified case was limited (~1200 -> 800 microseconds). However, the speed increase for the stratified case was impressive (~123 -> ~5 milliseconds)

Some lessons learned:

• For a single column, length(unique(x$col)) is noticeably faster than uniqueDT(x[, .(col)]), as least for small tables (nrow = 125 in this case). I assume that uniqueDT() pays off when the number of rows increases, and also when multiple columns need to be selected, but I didn't investigate thoroughly
• merge.data.table(x, y, by = "col", all.x = TRUE) matches the sort order of dplyr::left_join(x, y), whereas x[y, on = "col"] returns a different row order

[nanxstats]

Thanks for rewriting this, looking good to me.

Also, thanks for removing the magic numbers in tests - i think this will pay off in the long run.