samples.mean(), samples.std() etc. are not correctly ignoring nan values

[lukashergt]

Describe the bug
Simple summary statistics such as samples.mean() or samples.std() are not correctly computed when the data contains nan despite the skipna=True flag.

My suspicion is that this is happening when the data contains nan values, but the weights do not (e.g. when computing derived parameters that don't work for part of the parameter space).

To Reproduce

samples = read_chains("./tests/example_data/cb").remove_burn_in(burn_in=0.1)
samples['y'] = np.where(samples.index.get_level_values(0)<=2000, samples.x0 + samples.x1, np.nan)
print(samples.y.mean())
print(samples.y.mean(skipna=True))
print(np.ma.average(a=np.ma.array(samples.y, mask=np.isnan(samples.y)), weights=samples.get_weights()))
print(samples.loc[:2000].y.mean())
assert samples.y.mean() == samples.loc[:2000].y.mean()

Output:

-0.009548434286901122
-0.009548434286901122
-0.01623565975733643
-0.01623565975733643
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[29], line 7
      5 print(np.ma.average(a=np.ma.array(samples.y, mask=np.isnan(samples.y)), weights=samples.get_weights()))
      6 print(samples.loc[:2000].y.mean())
----> 7 assert samples.y.mean() == samples.loc[:2000].y.mean()

AssertionError:

Expected behavior
Any nan values should be completely ignored in the computation.

Additional context
This will also affect computation of things like the Gelman--Rubin statistic.