Add KDE stat

[mwaskom]

Largely based on the existing kdeplot API, although currently only supporting univariate KDE estimation (unlike in the functional interface, there will be separate objects for univariate / bivariate distribution transforms).

so.Plot(penguins, "flipper_length_mm").add(so.Area(), so.KDE())

One new trick is that with gridsize=None, it will evaluate the KDE at the original datapoints:

so.Plot(tips, "total_bill").add(so.Dots(), so.KDE(gridsize=None))

This will probably be more useful with a solution for #3053 so that you could have, e.g., a strip plot with color determined by density.

[codecov]

Codecov Report

Merging #3111 (13b1f55) into master (bf46954) will increase coverage by 0.01%.
The diff coverage is 100.00%.

❗ Current head 13b1f55 differs from pull request most recent head 21b2cee. Consider uploading reports for the commit 21b2cee to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3111      +/-   ##
==========================================
+ Coverage   98.39%   98.41%   +0.01%     
==========================================
  Files          74       76       +2     
  Lines       23810    24049     +239     
==========================================
+ Hits        23429    23668     +239     
  Misses        381      381              

Impacted Files	Coverage Δ
seaborn/_stats/base.py	96.15% <100.00%> (+2.03%)	⬆️
seaborn/_stats/counting.py	100.00% <100.00%> (ø)
seaborn/_stats/density.py	100.00% <100.00%> (ø)
seaborn/objects.py	100.00% <100.00%> (ø)
tests/_stats/test_counting.py	100.00% <100.00%> (ø)
tests/_stats/test_density.py	100.00% <100.00%> (ø)