[FEA] Support New Median / Median-Approximate in Dask-cuDF

[quasiben]

In dask/dask#9483 Dask now has an implementation of median and median_approximate. These should available with dask_cudf. cuDF currently raise a NotImplementedError with mean(axis=1):

cdf = cudf.datasets.timeseries()
cdf = dd.from_pandas(cdf, npartitions=2)
cdf.median(axis=1).compute()

File /datasets/bzaitlen/miniconda3/envs/tpcds-20220906/lib/python3.9/site-packages/cudf/core/dataframe.py:5225, in DataFrame.quantile(self, q, axis, numeric_only, interpolation, columns, exact)
   5164 """
   5165 Return values at the given quantile.
   5166
   (...)
   5222 0.5  2.5  55.0
   5223 """  # noqa: E501
   5224 if axis not in (0, None):
-> 5225     raise NotImplementedError("axis is not implemented yet")
   5227 data_df = self
   5228 if numeric_only:

As for mean_approximate, ValueError is thrown unexpectedly:

In [32]: cdf.compute()
Out[32]:
   x    y
0  1  1.1
1  2  2.2
2  3  3.3
3  4  4.4
4  5  5.5

In [33]: cdf.median_approximate().compute()
...
File /datasets/bzaitlen/miniconda3/envs/tpcds-20220906/lib/python3.9/site-packages/cudf/core/frame.py:898, in Frame.fillna(self, value, method, axis, inplace, limit)
    892 should_fill = (
    893     col_name in value
    894     and col.contains_na_entries
    895     and not libcudf.scalar._is_null_host_scalar(replace_val)
    896 ) or method is not None
    897 if should_fill:
--> 898     filled_data[col_name] = col.fillna(replace_val, method)
    899 else:
    900     filled_data[col_name] = col.copy(deep=True)

File /datasets/bzaitlen/miniconda3/envs/tpcds-20220906/lib/python3.9/site-packages/cudf/core/column/numerical.py:503, in NumericalColumn.fillna(self, fill_value, method, dtype, fill_nan)
    499     return super(NumericalColumn, col).fillna(fill_value, method)
    501 if np.isscalar(fill_value):
    502     # cast safely to the same dtype as self
--> 503     fill_value_casted = col.dtype.type(fill_value)
    504     if not np.isnan(fill_value) and (fill_value_casted != fill_value):
    505         raise TypeError(
    506             f"Cannot safely cast non-equivalent "
    507             f"{type(fill_value).__name__} to {col.dtype.name}"
    508         )

ValueError: cannot convert float NaN to integer