Fixed Diff primitive for rare nan case

[bschreck]

I ran into a bug with the Diff primitive that occurs when for a single variable and cutoff time, all values of both the variable and the feature to group by are nan. The bug occurs because Pandas drops the column names of empty dataframes, so we get a KeyError. This fix required a new variable in the mock entityset because we needed both an Id variable to group on that had a nan value and a numeric variable that had a nan.

[codecov-io]

Codecov Report

Merging #21 into master will increase coverage by 0.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master      #21      +/-   ##
==========================================
+ Coverage   86.83%   86.85%   +0.01%     
==========================================
  Files          73       73              
  Lines        6768     6777       +9     
==========================================
+ Hits         5877     5886       +9     
  Misses        891      891

Impacted Files	Coverage Δ
featuretools/tests/testing_utils/mock_ds.py	94.54% <100%> (ø)	⬆️
.../feature_function_tests/test_transform_features.py	85.3% <100%> (+0.15%)	⬆️
featuretools/primitives/transform_primitive.py	96.83% <100%> (+0.04%)	⬆️

---

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 fb87937...71e1571. Read the comment docs.

[kmax12]

Looks good to me