[FIX] appveyor: Bump scipy to 1.2.0

[thocevar]

Issue

Scipy.stats.mode handles nans incorrectly in version 1.0 (see #3480 (comment))

Description of changes

Bump scipy version to 1.2.0 on appveyor.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #3537 into master will increase coverage by <.01%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #3537      +/-   ##
==========================================
+ Coverage   83.67%   83.67%   +<.01%     
==========================================
  Files         370      370              
  Lines       66176    66176              
==========================================
+ Hits        55372    55373       +1     
+ Misses      10804    10803       -1

[janezd]

I'm reassigning this to @ales-erjavec. As I understand, it bumps the version in Appveyor test environment, but not in the build. Ales, can you verify that we use SciPy >= 1.2.0 in Windows builds? Also, it seems to me that we use numpy 1.9.3 on Windows, while we test on 1.14? Can you check?

Background: this PR is needed to fix a problem in #3480, which is due to a bug in scipy.stats.mode that is fixed in 1.2.0.

[ales-erjavec]

conda does not yet have scipy 1.2.0 builds for windows. The current miniconda based installer still has scipy 1.0.1 (although 1.1.0 packages are available).

Also, it seems to me that we use numpy 1.9.3 on Windows, while we test on 1.14? Can you check?

We build against numpy 1.9 for wider binary compatibility. The tests then run with 1.14. This is also the version in the installers.

[thocevar]

Unfortunately, scipy 1.1.0 has the same problem as 1.0.

[pavlin-policar]

Perhaps it would then be best if we wrap the mode function in statistics.utils and make it do what we want, and eventually remove it when it isn't needed anymore.