[FEA] Use _get_numeric_types in numeric reductions

[mrocklin]

If would be great if the following worked:

import cudf
df = cudf.datasets.timeseries()
df.mean()

Currently we get this

AttributeError: 'CategoricalColumn' object has no attribute 'mean'

Typically the solution in Pandas is to strip out the non-numeric columns first

df._get_numeric_data().mean()

We should consider adding this method call in all dataframe reductions that expect numeric data like sum, mean, prod, std, var, cum* and so on.

[mrocklin]

So to be fully explicit we would want changes like the following:

def mean(self, *kwargs):
-    return self._apply_support_method('mean', **kwargs)
+    return self._get_numeric_data()._apply_support_method('mean', **kwargs)

Along with tests that this works when we have a small dataframe with both numeric and text dtypes.

[beckernick]

I believe the behavior requested in this issue is now considered deprecated by pandas based on the following example:

import cudf
df = cudf.datasets.timeseries()
pdf = df.to_pandas()
pdf.mean()
/tmp/ipykernel_47228/2065928329.py:4: FutureWarning: Dropping of nuisance columns in DataFrame reductions (with 'numeric_only=None') is deprecated; in a future version this will raise TypeError.  Select only valid columns before calling the reduction.
  pdf.mean()
id    999.979244
x       0.000081
y       0.000099
dtype: float64

df.mean()
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
/tmp/ipykernel_47228/3698961737.py in <module>
----> 1 df.mean()

~/conda/envs/rapids-22.02/lib/python3.8/site-packages/cudf/core/frame.py in mean(self, axis, skipna, level, numeric_only, **kwargs)
   4304         dtype: float64
   4305         """
-> 4306         return self._reduce(
   4307             "mean",
   4308             axis=axis,

~/conda/envs/rapids-22.02/lib/python3.8/site-packages/cudf/core/dataframe.py in _reduce(self, op, axis, level, numeric_only, **kwargs)
   5326 
   5327         if axis == 0:
-> 5328             result = [
   5329                 getattr(self._data[col], op)(**kwargs)
   5330                 for col in self._data.names

~/conda/envs/rapids-22.02/lib/python3.8/site-packages/cudf/core/dataframe.py in <listcomp>(.0)
   5327         if axis == 0:
   5328             result = [
-> 5329                 getattr(self._data[col], op)(**kwargs)
   5330                 for col in self._data.names
   5331             ]

~/conda/envs/rapids-22.02/lib/python3.8/site-packages/cudf/core/column/column.py in mean(self, skipna, dtype)
   1178 
   1179     def mean(self, skipna: bool = None, dtype: Dtype = None):
-> 1180         raise TypeError(f"cannot perform mean with type {self.dtype}")
   1181 
   1182     def std(self, skipna: bool = None, ddof=1, dtype: Dtype = np.float64):

TypeError: cannot perform mean with type category

This is consistent with pandas-dev/pandas#41480 . Rather than match deprecated functionality, I think our current behavior of throwing a TypeError should be considered correct.

Separately, we have a bug gap with our numeric_only=True behavior, and I'll file an issue and cross-link.

[beckernick]

Rather than file a new issue, I'll add context here.

In Python, we provide passthrough support for the numeric_only parameter in most dataframe reductions for Dask compatibility. For DataFrame.rank, we implement the numeric_only parameter.

Implementing support for numeric_only in the DataFrame._reduce machinery would likely close this issue