BUG: Don't apply function to grouper

[TomAugspurger]

This came up in #5264, especially here, here, and this SO post.

Essentially, there are cases now where the agg or transform or apply function is applied to the grouper. This can cause failures in some very odd ways. As a simple example:

In [25]: df = pd.DataFrame({"A": ['a', 'b', 'a', 'b'], 'B': [1, 2, 3, 4]})

In [26]: df
Out[26]: 
   A  B
0  a  1
1  b  2
2  a  3
3  b  4

[4 rows x 2 columns]

In [28]: df.groupby('A').apply(lambda x: x+1)

TypeError: Could not operate 1 with block values cannot concatenate 'str' and 'int' objects

The example in the SO post linked above is especially subtle because the path the code follows depends on whether or not the aggfunc is successfully applied to the grouping column.

I'm going to poke around the code and see what breaks. I'd propose that we apply the function to the grouper in fewer (possibly zero) cases.

[jreback]

pushing to 0.15

[mroeschke]

This looks tested and fixed. Closing