Implement numeric_only for skew and kurtosis

[phofl]

• Closes #xxxx
• Tests added / passed
• Passes pre-commit run --all-files

cc @jrbourbeau

Not sure what the policy Is, but this changes behaviour for pandas 2.0, with this pr

df = pd.DataFrame(
    {
        "int": [1, 2, 3, 4, 5, 6, 7, 8],
        "dt": [pd.NaT] + [datetime(2010, i, 1) for i in range(1, 8)],
    }
)
ddf = dd.from_pandas(df, npartitions=2).skew()

raises like pandas does, but without this pr it drops the dt column. I guess? we would prefer not raising here without deprecating first since we did not show any deprecation warnings?

Nothing changes with pandas < 2.0.

[j-bennet]

Nice. Make sure to test upstream as well.

[jrbourbeau]

Thanks @phofl. Just one small comment

[jrbourbeau]

Can we add an assert_eq to make sure the result from dask matches pandas?

Also, is numeric_only=True tested elsewhere?

[phofl]

Both of them are tested somewhere else, that's why I added only the False case here.

[j-bennet]

I guess? we would prefer not raising here without deprecating first since we did not show any deprecation warnings?

I would rather raise. We should have added a deprecation earlier, that's true, but at least right now we have the chance to catch up with pandas' behavior. Supporting diverging behaviors and following one step behind feels like extra work for not much benefit. Just my 2c.

[phofl]

That's what we are doing, so all good