PERF: DataFrame.groupby.nunique is non-performant

[jreback]

xref #14376

# from the asv
In [10]: n = 10000
    ...:     df = DataFrame({'key1': randint(0, 500, size=n),
    ...:                              'key2': randint(0, 100, size=n),
    ...:                              'ints': randint(0, 1000, size=n),
    ...:                              'ints2': randint(0, 1000, size=n), })
    ...: 

In [11]: %timeit df.groupby(['key1', 'key2']).nunique()
1 loop, best of 3: 4.25 s per loop

In [12]: result = df.groupby(['key1', 'key2']).nunique()

In [13]: g = df.groupby(['key1', 'key2'])

In [14]: expected = pd.concat([getattr(g, col).nunique() for col in g._selected_obj.columns], axis=1)

In [15]: result.equals(expected)
Out[15]: True

In [16]: %timeit pd.concat([getattr(g, col).nunique() for col in g._selected_obj.columns], axis=1)
100 loops, best of 3: 6.94 ms per loop

Series.groupby.nunique has a very performant implementation, but the way the DataFrame.groupby.nunique is implemented (via .apply) it ends up in a python loop over the groups, which nullifies this.

should be straightforward to fix this. need to make sure to test with as_index=True/False

[jreback]

cc @xflr6