API: Consistently support `numeric_only` in groupby ops

[rhshadrach]

Inspecting the various groupby ops, I think the following are cases where we should have numeric_only.

• agg (API: GroupBy.agg() numeric_only deprecation with custom function #50538)
• aggregate (API: GroupBy.agg() numeric_only deprecation with custom function #50538)
• apply
• cumprod (ENH: Add numeric_only to DataFrame.cum* methods #53072)
• cumsum (ENH: Add numeric_only to DataFrame.cum* methods #53072)
• diff
• pct_change
• rank (ENH: No numeric_only argument for pandas.core.groupby.GroupBy.rank() #44438)
• transform
• ohlc

The following methods make sense to have numeric_only, but won't fail on any input (or hashable input for nunique and value_counts), and so I think it's okay if they don't. But is still nice to have.

• all
• any
• bfill
• count
• ffill
• nunique
• value_counts

The following methods should not get a numeric_only argument. They fall into a few typical camps: filters, plotting, or they do not depend on the columns (e.g. cumcount and size)

• boxplot
• cumcount
• describe # Handled by include="all"
• filter
• head
• hist
• nth
• pipe
• plot
• shift
• size
• tail
• take

[kwhkim]

Looking at the all possible methods that argument numeric_only= needs to support, I think...

wouldn't it be better to have independent method .numeric_only() to select columns with numeric values or number dtypes, since numeric_only= argument can conflict with the function to be .agg() in the parameter name which is numeric_only

[rhshadrach]

One can already do DataFrame.select_dtypes('numeric'). I think we should strive for consistency of arguments between DataFrame and groupby ops where it makes sense, and this is one of those cases. So unless we're going to deprecate numeric_only across the board, I'm still positive on including them in groupby.

[kwhkim]

But the purpose was to select numeric columns from DataFrameGroupby. I think .select_dtypes('numer') is applicable only on DataFrame object. Or we might add the method select_dtypes() also to DataFrameGroupyby.

I think for the purpose of dendency problem I would suggest to maintain it(numeric_only) and also get *arg, and **kwargs for the functions to be applied... (I mean in the case of .mean() or .std()... etc, because .mean(skipna=False) does not work...) When you think about it, numeric_only is not so popular parameter name and we can urge users to avoid it...(I wonder if there is any function that uses parameter name numeric_only or we can always make custom function with lambda x: f(x, numeric_only=True) or something similar.