[REVIEW] Correct performance regression with 0.7 groupby.

[thomcom]

An O(N^2) replacement operation was easily replicated with a cpp level replace. Thanks @shwina for your help!

Fixes #1685

[harrism]

Can you give us an idea of the performance difference between this, before this, and 0.6?

[thomcom]

This was the performance on branch-0.7 without the optimization:

https://www.dropbox.com/s/hvjfmifulpgmeny/Screenshot%202019-05-08%2022.55.56.png?raw=1

This is the performance now with the optimization:

https://www.dropbox.com/s/gawyiy5sd7dp0wy/Screenshot%202019-05-08%2023.04.40.png?raw=1

I can't check branch-0.6 at this time because I have other branches that I'm updating. Moving back to branch-0.6 could put me in a build unstable state for hours.

[jrhemstad]

It's wild that this could be done in 3 lines of code.

[harrism]

100x, not bad.