RemoveNaNColumns - compute nans in advance

[noahnovsak]

Current implementation works fine for numpy, but because dask is lazy it will iterate over the entire table multiple times (once for each attribute).
On a 3000 by 3000 example dataset this reduced the performance of logistic regression by a factor of 10. Without preprocessing it runs in ~500 msec, with preprocessing it took ~5 sec.

solution: compute nans once in advance and store them in memory.

[codecov]

Codecov Report

Merging #6532 (e0a3988) into dask (c1aa47f) will decrease coverage by 0.01%.
The diff coverage is 100.00%.

Additional details and impacted files

@@            Coverage Diff             @@
##             dask    #6532      +/-   ##
==========================================
- Coverage   87.70%   87.70%   -0.01%     
==========================================
  Files         322      322              
  Lines       70045    70045              
==========================================
- Hits        61432    61430       -2     
- Misses       8613     8615       +2