Speedup groupby transform calculations

[kmax12]

This PR updates the logic for how we calculate group by transform features.

Previously we iterated over the groups first and then the features, updating the resulting frame greedily as we went along.

Now we iterate over the features, then groups. For each feature, we accumate all the values across all groups and then update the frame just once.

Basically this improves the number of update calls from num features x num groups to just num features

Also added a benchmarks folder to hold the code I used to test these changes

[codecov]

Codecov Report

Merging #609 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #609      +/-   ##
==========================================
+ Coverage   97.42%   97.42%   +<.01%     
==========================================
  Files         118      118              
  Lines        9526     9532       +6     
==========================================
+ Hits         9281     9287       +6     
  Misses        245      245

Impacted Files	Coverage Δ
...s/computational_backends/feature_set_calculator.py	98.1% <100%> (+0.03%)	⬆️
featuretools/__init__.py	66.66% <100%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f6ebadc...e690733. Read the comment docs.

[rwedge]

I think the code looks good. Jupyter notebooks can be tricky to version control so I was wondering what approach you wanted to take for the benchmarks folder? When would we commit changes to the notebook?

[kmax12]

For now, what if we save notebooks with a cleared output? That will make them easier to diff. Just updated the current benchmark notebook with a cleared output

[rwedge]

That sounds reasonable